diff --git "a/parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_middle.json" "b/parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_middle.json" @@ -0,0 +1,39657 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 379, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 381, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 381, + 98 + ], + "score": 1.0, + "content": "HUMAN MOTION DIFFUSION MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 114, + 344, + 137 + ], + "lines": [ + { + "bbox": [ + 112, + 115, + 345, + 128 + ], + "spans": [ + { + "bbox": [ + 112, + 115, + 345, + 128 + ], + "score": 1.0, + "content": "Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 126, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 112, + 126, + 286, + 138 + ], + "score": 1.0, + "content": "Daniel Cohen-Or and Amit H. Bermano", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 113, + 137, + 250, + 159 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 219, + 149 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 219, + 149 + ], + "score": 1.0, + "content": "Tel Aviv University, Israel", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 149, + 252, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 149, + 252, + 159 + ], + "score": 1.0, + "content": "guytevet@mail.tau.ac.il", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 278, + 188, + 333, + 200 + ], + "lines": [ + { + "bbox": [ + 277, + 187, + 335, + 201 + ], + "spans": [ + { + "bbox": [ + 277, + 187, + 335, + 201 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 410 + ], + "lines": [ + { + "bbox": [ + 142, + 213, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 469, + 225 + ], + "score": 1.0, + "content": "Natural and expressive human motion generation is the holy grail of computer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "score": 1.0, + "content": "animation. It is a challenging task, due to the diversity of possible motion, hu-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 234, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 247 + ], + "score": 1.0, + "content": "man perceptual sensitivity to it, and the difficulty of accurately describing it.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "Therefore, current generative solutions are either low-quality or limited in ex-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "score": 1.0, + "content": "pressiveness. Diffusion models, which have already shown remarkable gener-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "score": 1.0, + "content": "ative capabilities in other domains, are promising candidates for human mo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "score": 1.0, + "content": "tion due to their many-to-many nature, but they tend to be resource hungry and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "hard to control. In this paper, we introduce Motion Diffusion Model (MDM),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 300, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 313 + ], + "score": 1.0, + "content": "a carefully adapted classifier-free diffusion-based generative model for the hu-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 311, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 324 + ], + "score": 1.0, + "content": "man motion domain. MDM is transformer-based, combining insights from mo-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "tion generation literature. A notable design-choice is the prediction of the sam-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "ple, rather than the noise, in each diffusion step. This facilitates the use of es-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "score": 1.0, + "content": "tablished geometric losses on the locations and velocities of the motion, such", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "score": 1.0, + "content": "as the foot contact loss. As we demonstrate, MDM is a generic approach, en-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "abling different modes of conditioning, and different generation tasks. We show", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "score": 1.0, + "content": "that our model is trained with lightweight resources and yet achieves state-of-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "score": 1.0, + "content": "the-art results on leading benchmarks for text-to-motion and action-to-motion 1.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 399, + 363, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 363, + 411 + ], + "score": 1.0, + "content": "https://guytevet.github.io/mdm-page/.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 430, + 206, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 208, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 208, + 444 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 449, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 464 + ], + "score": 1.0, + "content": "Human motion generation is a fundamental task in computer animation, with applications spanning", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "from gaming to robotics. It is a challenging field, due to several reasons, including the vast span of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 470, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 486 + ], + "score": 1.0, + "content": "possible motions, and the difficulty and cost of acquiring high quality data. For the recently emerging", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "text-to-motion setting, where motion is generated from natural language, another inherent problem", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "is data labeling. For example, the label ”kick��� could refer to a soccer kick, as well as a Karate one.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 504, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 517 + ], + "score": 1.0, + "content": "At the same time, given a specific kick there are many ways to describe it, from how it is performed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "to the emotions it conveys, constituting a many-to-many problem. Current approaches have shown", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "success in the field, demonstrating plausible mapping from text to motion (Petrovich et al., 2022;", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "Tevet et al., 2022; Ahuja & Morency, 2019). All these approaches, however, still limit the learned", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "distribution since they mainly employ auto-encoders or VAEs (Kingma & Welling, 2013) (implying", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "a one-to-one mapping or a normal latent distribution respectively). In this aspect, diffusion models", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "are a better candidate for human motion generation, as they are free from assumptions on the target", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "distribution, and are known for expressing well the many-to-many distribution matching problem", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 592, + 184, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 184, + 603 + ], + "score": 1.0, + "content": "we have described.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020; Ho et al., 2020) are a generative", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "approach that is gaining significant attention in the computer vision and graphics community. When", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "trained for conditioned generation, recent diffusion models (Ramesh et al., 2022; Saharia et al.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "2022b) have shown breakthroughs in terms of image quality and semantics. The competence of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 653, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 665 + ], + "score": 1.0, + "content": "these models have also been shown for other domains, including videos (Ho et al., 2022), and 3D", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "point clouds (Luo & Hu, 2021). The problem with such models, however, is that they are notoriously", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 675, + 299, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 299, + 687 + ], + "score": 1.0, + "content": "resource demanding and challenging to control.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 109, + 691, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "In this paper, we introduce Motion Diffusion Model (MDM) — a carefully adapted diffusion based", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "score": 1.0, + "content": "generative model for the human motion domain. Being diffusion-based, MDM gains from the na-", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 116, + 722, + 474, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 476, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 476, + 733 + ], + "score": 1.0, + "content": "1Code can be found at https://github.com/GuyTevet/motion-diffusion-model.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 379, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 381, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 381, + 98 + ], + "score": 1.0, + "content": "HUMAN MOTION DIFFUSION MODEL", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 114, + 344, + 137 + ], + "lines": [ + { + "bbox": [ + 112, + 115, + 345, + 128 + ], + "spans": [ + { + "bbox": [ + 112, + 115, + 345, + 128 + ], + "score": 1.0, + "content": "Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 126, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 112, + 126, + 286, + 138 + ], + "score": 1.0, + "content": "Daniel Cohen-Or and Amit H. Bermano", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 112, + 115, + 345, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 137, + 250, + 159 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 219, + 149 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 219, + 149 + ], + "score": 1.0, + "content": "Tel Aviv University, Israel", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 149, + 252, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 149, + 252, + 159 + ], + "score": 1.0, + "content": "guytevet@mail.tau.ac.il", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 112, + 136, + 252, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 188, + 333, + 200 + ], + "lines": [ + { + "bbox": [ + 277, + 187, + 335, + 201 + ], + "spans": [ + { + "bbox": [ + 277, + 187, + 335, + 201 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 410 + ], + "lines": [ + { + "bbox": [ + 142, + 213, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 469, + 225 + ], + "score": 1.0, + "content": "Natural and expressive human motion generation is the holy grail of computer", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 223, + 469, + 235 + ], + "score": 1.0, + "content": "animation. It is a challenging task, due to the diversity of possible motion, hu-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 234, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 469, + 247 + ], + "score": 1.0, + "content": "man perceptual sensitivity to it, and the difficulty of accurately describing it.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 142, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "Therefore, current generative solutions are either low-quality or limited in ex-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 268 + ], + "score": 1.0, + "content": "pressiveness. Diffusion models, which have already shown remarkable gener-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "score": 1.0, + "content": "ative capabilities in other domains, are promising candidates for human mo-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "score": 1.0, + "content": "tion due to their many-to-many nature, but they tend to be resource hungry and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "hard to control. In this paper, we introduce Motion Diffusion Model (MDM),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 300, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 469, + 313 + ], + "score": 1.0, + "content": "a carefully adapted classifier-free diffusion-based generative model for the hu-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 311, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 324 + ], + "score": 1.0, + "content": "man motion domain. MDM is transformer-based, combining insights from mo-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "tion generation literature. A notable design-choice is the prediction of the sam-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "ple, rather than the noise, in each diffusion step. This facilitates the use of es-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 142, + 344, + 469, + 355 + ], + "score": 1.0, + "content": "tablished geometric losses on the locations and velocities of the motion, such", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 469, + 367 + ], + "score": 1.0, + "content": "as the foot contact loss. As we demonstrate, MDM is a generic approach, en-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "abling different modes of conditioning, and different generation tasks. We show", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 469, + 389 + ], + "score": 1.0, + "content": "that our model is trained with lightweight resources and yet achieves state-of-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 469, + 399 + ], + "score": 1.0, + "content": "the-art results on leading benchmarks for text-to-motion and action-to-motion 1.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 399, + 363, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 363, + 411 + ], + "score": 1.0, + "content": "https://guytevet.github.io/mdm-page/.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14.5, + "bbox_fs": [ + 141, + 213, + 470, + 411 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 430, + 206, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 208, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 208, + 444 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 449, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 506, + 464 + ], + "score": 1.0, + "content": "Human motion generation is a fundamental task in computer animation, with applications spanning", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "from gaming to robotics. It is a challenging field, due to several reasons, including the vast span of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 470, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 486 + ], + "score": 1.0, + "content": "possible motions, and the difficulty and cost of acquiring high quality data. For the recently emerging", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "text-to-motion setting, where motion is generated from natural language, another inherent problem", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "is data labeling. For example, the label ”kick” could refer to a soccer kick, as well as a Karate one.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 504, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 517 + ], + "score": 1.0, + "content": "At the same time, given a specific kick there are many ways to describe it, from how it is performed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "to the emotions it conveys, constituting a many-to-many problem. Current approaches have shown", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 538 + ], + "score": 1.0, + "content": "success in the field, demonstrating plausible mapping from text to motion (Petrovich et al., 2022;", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "Tevet et al., 2022; Ahuja & Morency, 2019). All these approaches, however, still limit the learned", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 506, + 561 + ], + "score": 1.0, + "content": "distribution since they mainly employ auto-encoders or VAEs (Kingma & Welling, 2013) (implying", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "a one-to-one mapping or a normal latent distribution respectively). In this aspect, diffusion models", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "are a better candidate for human motion generation, as they are free from assumptions on the target", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "distribution, and are known for expressing well the many-to-many distribution matching problem", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 592, + 184, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 184, + 603 + ], + "score": 1.0, + "content": "we have described.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 447, + 506, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020; Ho et al., 2020) are a generative", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "approach that is gaining significant attention in the computer vision and graphics community. When", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "trained for conditioned generation, recent diffusion models (Ramesh et al., 2022; Saharia et al.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "score": 1.0, + "content": "2022b) have shown breakthroughs in terms of image quality and semantics. The competence of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 653, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 665 + ], + "score": 1.0, + "content": "these models have also been shown for other domains, including videos (Ho et al., 2022), and 3D", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "point clouds (Luo & Hu, 2021). The problem with such models, however, is that they are notoriously", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 675, + 299, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 299, + 687 + ], + "score": 1.0, + "content": "resource demanding and challenging to control.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 609, + 506, + 687 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 691, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "In this paper, we introduce Motion Diffusion Model (MDM) — a carefully adapted diffusion based", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 716 + ], + "score": 1.0, + "content": "generative model for the human motion domain. Being diffusion-based, MDM gains from the na-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 208, + 343, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 343, + 220 + ], + "score": 1.0, + "content": "“A man runs to the right then runs to the left then back to the middle.”", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 46.5, + "bbox_fs": [ + 106, + 691, + 505, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 80, + 490, + 201 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 80, + 490, + 201 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 81, + 490, + 201 + ], + "spans": [ + { + "bbox": [ + 112, + 81, + 490, + 201 + ], + "score": 0.848, + "type": "image", + "image_path": "679e88c7d9a9e84834db9606fa91131fa59019df65a3ddf03e3c9875429bf51b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 80, + 490, + 120.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 120.33333333333334, + 490, + 160.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 160.66666666666669, + 490, + 201.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 209, + 342, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 343, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 343, + 220 + ], + "score": 1.0, + "content": "“A man runs to the right then runs to the left then back to the middle.”", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 107, + 223, + 505, + 321 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 223, + 505, + 321 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 223, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 107, + 223, + 505, + 321 + ], + "score": 0.957, + "type": "image", + "image_path": "8f0257ea475bbecd824964d2530d3df4475129bcbbba57fb1a6abf92a0e2e30c.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 223, + 505, + 255.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 255.66666666666666, + 505, + 288.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 288.3333333333333, + 505, + 321.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 506, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "Figure 1: Our Motion Diffusion Model (MDM) reflects the many-to-many nature of text-to-motion", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "mapping by generating diverse motions given a text prompt. Our custom architecture and geometric", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 355, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 479, + 369 + ], + "score": 1.0, + "content": "losses help yielding high-quality motion. Darker color indicates later frames in the sequence.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 108, + 379, + 502, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "score": 1.0, + "content": "tive aforementioned many-to-many expression of the domain, as evidenced by the resulting motion", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "quality and diversity (Figure 1). In addition, MDM combines insights already well established in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 401, + 475, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 475, + 414 + ], + "score": 1.0, + "content": "the motion generation domain, helping it be significantly more lightweight and controllable.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "First, instead of the ubiquitous U-net (Ronneberger et al., 2015) backbone, MDM is transformer-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "based. As we demonstrate, our architecture (Figure 2) is lightweight and better fits the temporal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "score": 1.0, + "content": "and spatially irregular nature of motion data (represented as a collection of joints). A large vol-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "ume of motion generation research is devoted to learning using geometric losses (Kocabas et al.,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "2020; Harvey et al., 2020; Aberman et al., 2020). Some, for example, regulate the velocity of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "motion (Petrovich et al., 2021) to prevent jitter, or specifically consider foot sliding using dedicated", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "terms (Shi et al., 2020). Consistently with these works, we show that applying geometric losses in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 495, + 275, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 275, + 507 + ], + "score": 1.0, + "content": "the diffusion setting improves generation.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "The MDM framework has a generic design enabling different forms of conditioning. We showcase", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "three tasks: text-to-motion, action-to-motion, and unconditioned generation. We train the model", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "score": 1.0, + "content": "in a classifier-free manner (Ho & Salimans, 2022), which enables trading-off diversity to fidelity,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "and sampling both conditionally and unconditionally from the same model. In the text-to-motion", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "task, our model generates coherent motions (Figure 1) that achieve state-of-the-art results on the Hu-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 565, + 504, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 504, + 578 + ], + "score": 1.0, + "content": "manML3D (Guo et al., 2022a) and KIT (Plappert et al., 2016) benchmarks. Moreover, our user study", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 408, + 590 + ], + "score": 1.0, + "content": "shows that human evaluators prefer our generated motions over real motions", + "type": "text" + }, + { + "bbox": [ + 408, + 577, + 428, + 588 + ], + "score": 0.89, + "content": "4 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "of the time (Figure", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "4(a)). In action-to-motion, MDM outperforms the state-of-the-art (Guo et al., 2020; Petrovich et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "2021), even though they were specifically designed for this task, on the common HumanAct12 (Guo", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 609, + 323, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 323, + 622 + ], + "score": 1.0, + "content": "et al., 2020) and UESTC (Ji et al., 2018) benchmarks.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 104, + 624, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 506, + 641 + ], + "score": 1.0, + "content": "Lastly, we also demonstrate completion and editing. By adapting diffusion image-inpainting (Song", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "et al., 2020b; Saharia et al., 2022a), we set a motion prefix and suffix, and use our model to fill in the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "gap. Doing so under a textual condition guides MDM to fill the gap with a specific motion that still", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "maintains the semantics of the original input. By performing inpainting in the joints space rather", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "than temporally, we also demonstrate the semantic editing of specific body parts, without changing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 193, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 193, + 694 + ], + "score": 1.0, + "content": "the others (Figure 3).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Overall, we introduce Motion Diffusion Model, a motion framework that achieves state-of-the-art", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "quality in several motion generation tasks, while requiring only about three days of training on a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "single mid-range GPU. It supports geometric losses, which are non trivial to the diffusion setting,", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 80, + 490, + 201 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 80, + 490, + 201 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 81, + 490, + 201 + ], + "spans": [ + { + "bbox": [ + 112, + 81, + 490, + 201 + ], + "score": 0.848, + "type": "image", + "image_path": "679e88c7d9a9e84834db9606fa91131fa59019df65a3ddf03e3c9875429bf51b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 109, + 80, + 490, + 120.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 120.33333333333334, + 490, + 160.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 160.66666666666669, + 490, + 201.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 209, + 342, + 219 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 106, + 208, + 343, + 220 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 107, + 223, + 505, + 321 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 223, + 505, + 321 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 223, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 107, + 223, + 505, + 321 + ], + "score": 0.957, + "type": "image", + "image_path": "8f0257ea475bbecd824964d2530d3df4475129bcbbba57fb1a6abf92a0e2e30c.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 223, + 505, + 255.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 255.66666666666666, + 505, + 288.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 288.3333333333333, + 505, + 321.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 334, + 506, + 367 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "Figure 1: Our Motion Diffusion Model (MDM) reflects the many-to-many nature of text-to-motion", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "mapping by generating diverse motions given a text prompt. Our custom architecture and geometric", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 355, + 479, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 479, + 369 + ], + "score": 1.0, + "content": "losses help yielding high-quality motion. Darker color indicates later frames in the sequence.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 108, + 379, + 502, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "score": 1.0, + "content": "tive aforementioned many-to-many expression of the domain, as evidenced by the resulting motion", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "quality and diversity (Figure 1). In addition, MDM combines insights already well established in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 401, + 475, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 475, + 414 + ], + "score": 1.0, + "content": "the motion generation domain, helping it be significantly more lightweight and controllable.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 378, + 505, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 417, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "First, instead of the ubiquitous U-net (Ronneberger et al., 2015) backbone, MDM is transformer-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "based. As we demonstrate, our architecture (Figure 2) is lightweight and better fits the temporal", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 451 + ], + "score": 1.0, + "content": "and spatially irregular nature of motion data (represented as a collection of joints). A large vol-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "ume of motion generation research is devoted to learning using geometric losses (Kocabas et al.,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "2020; Harvey et al., 2020; Aberman et al., 2020). Some, for example, regulate the velocity of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "motion (Petrovich et al., 2021) to prevent jitter, or specifically consider foot sliding using dedicated", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "terms (Shi et al., 2020). Consistently with these works, we show that applying geometric losses in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 495, + 275, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 275, + 507 + ], + "score": 1.0, + "content": "the diffusion setting improves generation.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 418, + 506, + 507 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "The MDM framework has a generic design enabling different forms of conditioning. We showcase", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "three tasks: text-to-motion, action-to-motion, and unconditioned generation. We train the model", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "score": 1.0, + "content": "in a classifier-free manner (Ho & Salimans, 2022), which enables trading-off diversity to fidelity,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "and sampling both conditionally and unconditionally from the same model. In the text-to-motion", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "task, our model generates coherent motions (Figure 1) that achieve state-of-the-art results on the Hu-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 565, + 504, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 504, + 578 + ], + "score": 1.0, + "content": "manML3D (Guo et al., 2022a) and KIT (Plappert et al., 2016) benchmarks. Moreover, our user study", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 577, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 408, + 590 + ], + "score": 1.0, + "content": "shows that human evaluators prefer our generated motions over real motions", + "type": "text" + }, + { + "bbox": [ + 408, + 577, + 428, + 588 + ], + "score": 0.89, + "content": "4 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 577, + 505, + 590 + ], + "score": 1.0, + "content": "of the time (Figure", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "4(a)). In action-to-motion, MDM outperforms the state-of-the-art (Guo et al., 2020; Petrovich et al.,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 612 + ], + "score": 1.0, + "content": "2021), even though they were specifically designed for this task, on the common HumanAct12 (Guo", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 609, + 323, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 323, + 622 + ], + "score": 1.0, + "content": "et al., 2020) and UESTC (Ji et al., 2018) benchmarks.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 511, + 506, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 104, + 624, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 506, + 641 + ], + "score": 1.0, + "content": "Lastly, we also demonstrate completion and editing. By adapting diffusion image-inpainting (Song", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "et al., 2020b; Saharia et al., 2022a), we set a motion prefix and suffix, and use our model to fill in the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 662 + ], + "score": 1.0, + "content": "gap. Doing so under a textual condition guides MDM to fill the gap with a specific motion that still", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "maintains the semantics of the original input. By performing inpainting in the joints space rather", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "than temporally, we also demonstrate the semantic editing of specific body parts, without changing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 193, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 193, + 694 + ], + "score": 1.0, + "content": "the others (Figure 3).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 624, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Overall, we introduce Motion Diffusion Model, a motion framework that achieves state-of-the-art", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "quality in several motion generation tasks, while requiring only about three days of training on a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "single mid-range GPU. It supports geometric losses, which are non trivial to the diffusion setting,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "but are crucial to the motion domain, and offers the combination of state-of-the-art generative power", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 276, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 276, + 106 + ], + "score": 1.0, + "content": "with well thought-out domain knowledge.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 699, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "but are crucial to the motion domain, and offers the combination of state-of-the-art generative power", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 276, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 276, + 106 + ], + "score": 1.0, + "content": "with well thought-out domain knowledge.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 108, + 116, + 211, + 129 + ], + "lines": [ + { + "bbox": [ + 104, + 115, + 213, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 213, + 131 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 108, + 136, + 263, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 264, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 264, + 149 + ], + "score": 1.0, + "content": "2.1 HUMAN MOTION GENERATION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "score": 1.0, + "content": "Neural motion generation, learned from motion capture data, can be conditioned by any signal that", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "describes the motion. Many works use parts of the motion itself for guidance. Some predict motion", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "from its prefix poses (Fragkiadaki et al., 2015; Martinez et al., 2017; Hernandez et al., 2019; Guo", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "et al., 2022b). Others (Harvey & Pal, 2018; Kaufmann et al., 2020; Harvey et al., 2020; Duan et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "2021) solve in-betweening and super-resolution tasks using bi-directional GRU (Cho et al., 2014)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "score": 1.0, + "content": "and Transformer (Vaswani et al., 2017) architectures. Holden et al. (2016) use auto-encoder to learn", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "motion latent representation, then utilize it to edit and control motion with spatial constraints such", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "as root trajectory and bone lengths. Motion can be controlled with a high-level guidance given from", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "action class (Guo et al., 2020; Petrovich et al., 2021; Cervantes et al., 2022), audio (Li et al., 2021;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Aristidou et al., 2022) and natural language (Ahuja & Morency, 2019; Petrovich et al., 2022). In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 493, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 493, + 274 + ], + "score": 1.0, + "content": "most cases authors suggests a dedicated approach to map each conditioning domain into motion.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "In recent years, the leading approach for the Text-to-Motion task is to learn a shared latent space", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "score": 1.0, + "content": "for language and motion. JL2P (Ahuja & Morency, 2019) learns the KIT motion-language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 300, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 314 + ], + "score": 1.0, + "content": "dataset (Plappert et al., 2016) with an auto-encoder, limiting one-to-one mapping from text to mo-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "score": 1.0, + "content": "tion. TEMOS (Petrovich et al., 2022) and T2M (Guo et al., 2022a) suggest using a VAE (Kingma", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "& Welling, 2013) to map a text prompt into a normal distribution in latent space. Recently, Mo-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "tionCLIP (Tevet et al., 2022) leverages the shared text-image latent space learned by CLIP (Radford", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 495, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 495, + 358 + ], + "score": 1.0, + "content": "et al., 2021) to expand text-to-motion out of the data limitations and enabled latent space editing.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 361, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "The human motion manifold can also be learned without labels, as shown by Holden et al. (2016),", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "V-Poser (Pavlakos et al., 2019), and more recently the dedicated MoDi architecture (Raab et al.,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 383, + 440, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 440, + 396 + ], + "score": 1.0, + "content": "2022). We show that our model is capable for such an unsupervised setting as well.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 109, + 403, + 277, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 278, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 278, + 416 + ], + "score": 1.0, + "content": "2.2 DIFFUSION GENERATIVE MODELS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 420, + 505, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020) are a class of neural generative", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 430, + 504, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 504, + 442 + ], + "score": 1.0, + "content": "models, based on the stochastic diffusion process as it is modeled in Thermodynamics. In this set-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "ting, a sample from the data distribution is gradually noised by the diffusion process. Then, a neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "model learns the reverse process of gradually denoising the sample. Sampling the learned data dis-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "tribution is done by denoising a pure initial noise. Ho et al. (2020) and Song et al. (2020a) further", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "developed the practices for image generation applications. For conditioned generation, Dhariwal &", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "Nichol (2021), introduced classifier-guided diffusion, which was later on adapted by GLIDE (Nichol", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "et al., 2021) to enable conditioning over CLIP textual representations. The Classifier-Free Guidance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "approach Ho & Salimans (2022) enables conditioning while trading-off fidelity and diversity, and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "achieves better results (Nichol et al., 2021). In this paper, we implement text-to-motion by condi-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "tioning on CLIP in a classifier-free manner, similarly to text-to-image (Ramesh et al., 2022; Saharia", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "score": 1.0, + "content": "et al., 2022b). Local editing of images is typically defined as an inpainting problem, where a part", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "of the image is constant, and the inpainted part is denoised by the model, possibly under some con-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "dition (Song et al., 2020b; Saharia et al., 2022a). We adapt this technique to edit motion’s specific", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 573, + 444, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 444, + 586 + ], + "score": 1.0, + "content": "body parts or temporal intervals (in-betweening) according to an optional condition.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 504, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "Closer to our context, Gu et al. (2022) used the diffusion formulation to model the stochasticity of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "human trajectory prediction. More recently, concurrent to this work, Zhang et al. (2022) and Kim", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "et al. (2022) have suggested diffusion models for motion generation. Our work requires significantly", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "score": 1.0, + "content": "fewer GPU resources and makes design choices that enable geometric losses, which improve results.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 271, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 273, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 273, + 660 + ], + "score": 1.0, + "content": "3 MOTION DIFFUSION MODEL", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "An overview of our method is described in Figure 2. Our goal is to synthesize a human motion", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 107, + 673, + 507, + 691 + ], + "spans": [ + { + "bbox": [ + 107, + 676, + 127, + 687 + ], + "score": 0.88, + "content": "x ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 673, + 168, + 691 + ], + "score": 1.0, + "content": "of length", + "type": "text" + }, + { + "bbox": [ + 169, + 677, + 179, + 687 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 673, + 296, + 691 + ], + "score": 1.0, + "content": "given an arbitrary condition", + "type": "text" + }, + { + "bbox": [ + 297, + 679, + 302, + 687 + ], + "score": 0.6, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 673, + 507, + 691 + ], + "score": 1.0, + "content": ". This condition can be any real-world signal that", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 685, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 702 + ], + "score": 1.0, + "content": "will dictate the synthesis, such as audio (Li et al., 2021; Aristidou et al., 2022), natural language", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "(text-to-motion) (Tevet et al., 2022; Guo et al., 2022a) or a discrete class (action-to-motion) (Guo", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 102, + 710, + 509, + 738 + ], + "spans": [ + { + "bbox": [ + 102, + 710, + 260, + 738 + ], + "score": 1.0, + "content": "et al., 2020; Petrovich et al., 2021). Iwhich we denote as the null condition", + "type": "text" + }, + { + "bbox": [ + 261, + 721, + 284, + 731 + ], + "score": 0.89, + "content": "c = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 710, + 378, + 738 + ], + "score": 1.0, + "content": "tion, unconditioned mo. The generated motion", + "type": "text" + }, + { + "bbox": [ + 379, + 720, + 444, + 733 + ], + "score": 0.94, + "content": "\\boldsymbol { x } ^ { 1 : N } = \\{ x ^ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 710, + 509, + 738 + ], + "score": 1.0, + "content": "s also possible,is a sequences", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 116, + 211, + 129 + ], + "lines": [ + { + "bbox": [ + 104, + 115, + 213, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 213, + 131 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 108, + 136, + 263, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 136, + 264, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 136, + 264, + 149 + ], + "score": 1.0, + "content": "2.1 HUMAN MOTION GENERATION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 505, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 506, + 165 + ], + "score": 1.0, + "content": "Neural motion generation, learned from motion capture data, can be conditioned by any signal that", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "describes the motion. Many works use parts of the motion itself for guidance. Some predict motion", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "from its prefix poses (Fragkiadaki et al., 2015; Martinez et al., 2017; Hernandez et al., 2019; Guo", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "et al., 2022b). Others (Harvey & Pal, 2018; Kaufmann et al., 2020; Harvey et al., 2020; Duan et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "2021) solve in-betweening and super-resolution tasks using bi-directional GRU (Cho et al., 2014)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 219 + ], + "score": 1.0, + "content": "and Transformer (Vaswani et al., 2017) architectures. Holden et al. (2016) use auto-encoder to learn", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "motion latent representation, then utilize it to edit and control motion with spatial constraints such", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "as root trajectory and bone lengths. Motion can be controlled with a high-level guidance given from", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "action class (Guo et al., 2020; Petrovich et al., 2021; Cervantes et al., 2022), audio (Li et al., 2021;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 264 + ], + "score": 1.0, + "content": "Aristidou et al., 2022) and natural language (Ahuja & Morency, 2019; Petrovich et al., 2022). In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 493, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 493, + 274 + ], + "score": 1.0, + "content": "most cases authors suggests a dedicated approach to map each conditioning domain into motion.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 153, + 506, + 274 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 291 + ], + "score": 1.0, + "content": "In recent years, the leading approach for the Text-to-Motion task is to learn a shared latent space", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 303 + ], + "score": 1.0, + "content": "for language and motion. JL2P (Ahuja & Morency, 2019) learns the KIT motion-language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 300, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 505, + 314 + ], + "score": 1.0, + "content": "dataset (Plappert et al., 2016) with an auto-encoder, limiting one-to-one mapping from text to mo-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "score": 1.0, + "content": "tion. TEMOS (Petrovich et al., 2022) and T2M (Guo et al., 2022a) suggest using a VAE (Kingma", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 505, + 335 + ], + "score": 1.0, + "content": "& Welling, 2013) to map a text prompt into a normal distribution in latent space. Recently, Mo-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "tionCLIP (Tevet et al., 2022) leverages the shared text-image latent space learned by CLIP (Radford", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 343, + 495, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 495, + 358 + ], + "score": 1.0, + "content": "et al., 2021) to expand text-to-motion out of the data limitations and enabled latent space editing.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 279, + 506, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 361, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 374 + ], + "score": 1.0, + "content": "The human motion manifold can also be learned without labels, as shown by Holden et al. (2016),", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "V-Poser (Pavlakos et al., 2019), and more recently the dedicated MoDi architecture (Raab et al.,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 383, + 440, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 440, + 396 + ], + "score": 1.0, + "content": "2022). We show that our model is capable for such an unsupervised setting as well.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 360, + 505, + 396 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 403, + 277, + 415 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 278, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 278, + 416 + ], + "score": 1.0, + "content": "2.2 DIFFUSION GENERATIVE MODELS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 420, + 505, + 584 + ], + "lines": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020) are a class of neural generative", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 430, + 504, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 504, + 442 + ], + "score": 1.0, + "content": "models, based on the stochastic diffusion process as it is modeled in Thermodynamics. In this set-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 453 + ], + "score": 1.0, + "content": "ting, a sample from the data distribution is gradually noised by the diffusion process. Then, a neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "model learns the reverse process of gradually denoising the sample. Sampling the learned data dis-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "tribution is done by denoising a pure initial noise. Ho et al. (2020) and Song et al. (2020a) further", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "developed the practices for image generation applications. For conditioned generation, Dhariwal &", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 498 + ], + "score": 1.0, + "content": "Nichol (2021), introduced classifier-guided diffusion, which was later on adapted by GLIDE (Nichol", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "et al., 2021) to enable conditioning over CLIP textual representations. The Classifier-Free Guidance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "approach Ho & Salimans (2022) enables conditioning while trading-off fidelity and diversity, and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "achieves better results (Nichol et al., 2021). In this paper, we implement text-to-motion by condi-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "tioning on CLIP in a classifier-free manner, similarly to text-to-image (Ramesh et al., 2022; Saharia", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 553 + ], + "score": 1.0, + "content": "et al., 2022b). Local editing of images is typically defined as an inpainting problem, where a part", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "of the image is constant, and the inpainted part is denoised by the model, possibly under some con-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "dition (Song et al., 2020b; Saharia et al., 2022a). We adapt this technique to edit motion’s specific", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 573, + 444, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 444, + 586 + ], + "score": 1.0, + "content": "body parts or temporal intervals (in-betweening) according to an optional condition.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 419, + 506, + 586 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 504, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "Closer to our context, Gu et al. (2022) used the diffusion formulation to model the stochasticity of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "human trajectory prediction. More recently, concurrent to this work, Zhang et al. (2022) and Kim", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "et al. (2022) have suggested diffusion models for motion generation. Our work requires significantly", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 504, + 635 + ], + "score": 1.0, + "content": "fewer GPU resources and makes design choices that enable geometric losses, which improve results.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 590, + 506, + 635 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 271, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 273, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 273, + 660 + ], + "score": 1.0, + "content": "3 MOTION DIFFUSION MODEL", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "An overview of our method is described in Figure 2. Our goal is to synthesize a human motion", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 107, + 673, + 507, + 691 + ], + "spans": [ + { + "bbox": [ + 107, + 676, + 127, + 687 + ], + "score": 0.88, + "content": "x ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 673, + 168, + 691 + ], + "score": 1.0, + "content": "of length", + "type": "text" + }, + { + "bbox": [ + 169, + 677, + 179, + 687 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 673, + 296, + 691 + ], + "score": 1.0, + "content": "given an arbitrary condition", + "type": "text" + }, + { + "bbox": [ + 297, + 679, + 302, + 687 + ], + "score": 0.6, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 673, + 507, + 691 + ], + "score": 1.0, + "content": ". This condition can be any real-world signal that", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 685, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 702 + ], + "score": 1.0, + "content": "will dictate the synthesis, such as audio (Li et al., 2021; Aristidou et al., 2022), natural language", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "(text-to-motion) (Tevet et al., 2022; Guo et al., 2022a) or a discrete class (action-to-motion) (Guo", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 102, + 710, + 509, + 738 + ], + "spans": [ + { + "bbox": [ + 102, + 710, + 260, + 738 + ], + "score": 1.0, + "content": "et al., 2020; Petrovich et al., 2021). Iwhich we denote as the null condition", + "type": "text" + }, + { + "bbox": [ + 261, + 721, + 284, + 731 + ], + "score": 0.89, + "content": "c = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 710, + 378, + 738 + ], + "score": 1.0, + "content": "tion, unconditioned mo. The generated motion", + "type": "text" + }, + { + "bbox": [ + 379, + 720, + 444, + 733 + ], + "score": 0.94, + "content": "\\boldsymbol { x } ^ { 1 : N } = \\{ x ^ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 710, + 509, + 738 + ], + "score": 1.0, + "content": "s also possible,is a sequences", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 104, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 308, + 362, + 322 + ], + "score": 1.0, + "content": "of human poses represented by either joint rotations or positions", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 362, + 309, + 410, + 320 + ], + "score": 0.92, + "content": "\\boldsymbol { x } ^ { i } \\in \\mathbb { R } ^ { J \\times D }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 410, + 308, + 440, + 322 + ], + "score": 1.0, + "content": ", where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 441, + 310, + 448, + 320 + ], + "score": 0.84, + "content": "J", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 448, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "is the number", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 160, + 333 + ], + "score": 1.0, + "content": "of joints and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 160, + 321, + 170, + 331 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 171, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "is the dimension of the joint representation. MDM can accept motion represented", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 332, + 321, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 321, + 344 + ], + "score": 1.0, + "content": "by either locations, rotations, or both (see Section 4).", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 48, + "bbox_fs": [ + 102, + 665, + 509, + 738 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 75, + 504, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 75, + 504, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 75, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 75, + 504, + 213 + ], + "score": 0.969, + "type": "image", + "image_path": "aceed64cf2b4fbe7a9a4e3115da6b56c1525a7c001a69455abf0263ac36383e4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 75, + 504, + 121.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 121.0, + 504, + 167.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 167.0, + 504, + 213.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 220, + 505, + 298 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "Figure 2: (Left) Motion Diffusion Model (MDM) overview. The model is fed a motion sequence", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 226, + 509, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 127, + 244 + ], + "score": 0.91, + "content": "x _ { t } ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 226, + 165, + 248 + ], + "score": 1.0, + "content": "of length", + "type": "text" + }, + { + "bbox": [ + 166, + 232, + 176, + 242 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 226, + 243, + 248 + ], + "score": 1.0, + "content": "in a noising step", + "type": "text" + }, + { + "bbox": [ + 244, + 233, + 249, + 242 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 226, + 292, + 248 + ], + "score": 1.0, + "content": ", as well as", + "type": "text" + }, + { + "bbox": [ + 292, + 233, + 297, + 242 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 226, + 509, + 248 + ], + "score": 1.0, + "content": "itself and a conditioning code c. c, a CLIP (Radford", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 241, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 506, + 257 + ], + "score": 1.0, + "content": "et al., 2021) based textual embedding in this case, is first randomly masked for classifier-free learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 239, + 266 + ], + "score": 1.0, + "content": "and then projected together with", + "type": "text" + }, + { + "bbox": [ + 239, + 255, + 244, + 264 + ], + "score": 0.6, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 255, + 326, + 266 + ], + "score": 1.0, + "content": "into the input token", + "type": "text" + }, + { + "bbox": [ + 326, + 255, + 340, + 265 + ], + "score": 0.87, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 255, + 505, + 266 + ], + "score": 1.0, + "content": ". In each sampling step, the transformer-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 262, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 102, + 262, + 268, + 282 + ], + "score": 1.0, + "content": "encoder predicts the final clean motion", + "type": "text" + }, + { + "bbox": [ + 268, + 264, + 289, + 277 + ], + "score": 0.91, + "content": "\\hat { x } _ { 0 } ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 262, + 480, + 282 + ], + "score": 1.0, + "content": ". (Right) Sampling MDM. Given a condition", + "type": "text" + }, + { + "bbox": [ + 480, + 267, + 486, + 275 + ], + "score": 0.54, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 262, + 506, + 282 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 276, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 194, + 287 + ], + "score": 1.0, + "content": "sample random noise", + "type": "text" + }, + { + "bbox": [ + 194, + 277, + 207, + 287 + ], + "score": 0.86, + "content": "x _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 276, + 441, + 287 + ], + "score": 1.0, + "content": "at the dimensions of the desired motion, then iterate from", + "type": "text" + }, + { + "bbox": [ + 441, + 276, + 449, + 286 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 276, + 505, + 287 + ], + "score": 1.0, + "content": "to 1. At each", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 286, + 394, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 125, + 301 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 125, + 288, + 130, + 297 + ], + "score": 0.6, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 286, + 263, + 301 + ], + "score": 1.0, + "content": ", MDM predicts the clean sample", + "type": "text" + }, + { + "bbox": [ + 264, + 287, + 275, + 298 + ], + "score": 0.87, + "content": "{ \\hat { x } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 286, + 369, + 301 + ], + "score": 1.0, + "content": ", and diffuses it back to", + "type": "text" + }, + { + "bbox": [ + 370, + 289, + 389, + 298 + ], + "score": 0.89, + "content": "x _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 286, + 394, + 301 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 308, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 104, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 104, + 308, + 362, + 322 + ], + "score": 1.0, + "content": "of human poses represented by either joint rotations or positions", + "type": "text" + }, + { + "bbox": [ + 362, + 309, + 410, + 320 + ], + "score": 0.92, + "content": "\\boldsymbol { x } ^ { i } \\in \\mathbb { R } ^ { J \\times D }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 308, + 440, + 322 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 441, + 310, + 448, + 320 + ], + "score": 0.84, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "is the number", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 160, + 333 + ], + "score": 1.0, + "content": "of joints and", + "type": "text" + }, + { + "bbox": [ + 160, + 321, + 170, + 331 + ], + "score": 0.8, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "is the dimension of the joint representation. MDM can accept motion represented", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 332, + 321, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 321, + 344 + ], + "score": 1.0, + "content": "by either locations, rotations, or both (see Section 4).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 105, + 347, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 104, + 344, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 344, + 370, + 365 + ], + "score": 1.0, + "content": "Framework. Diffusion is modeled as a Markov noising process,", + "type": "text" + }, + { + "bbox": [ + 371, + 347, + 414, + 361 + ], + "score": 0.93, + "content": "\\{ x _ { t } ^ { 1 : N } \\} _ { t = 0 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 344, + 446, + 365 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 446, + 348, + 466, + 361 + ], + "score": 0.92, + "content": "x _ { 0 } ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 344, + 507, + 365 + ], + "score": 1.0, + "content": "is drawn", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 360, + 227, + 370 + ], + "spans": [ + { + "bbox": [ + 107, + 360, + 227, + 370 + ], + "score": 1.0, + "content": "from the data distribution and", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 374, + 391, + 390 + ], + "lines": [ + { + "bbox": [ + 219, + 374, + 391, + 390 + ], + "spans": [ + { + "bbox": [ + 219, + 374, + 391, + 390 + ], + "score": 0.93, + "content": "q ( x _ { t } ^ { 1 : N } | x _ { t - 1 } ^ { 1 : N } ) = \\mathcal { N } ( \\sqrt { \\alpha _ { t } } x _ { t - 1 } ^ { 1 : N } , ( 1 - \\alpha _ { t } ) I ) ,", + "type": "interline_equation", + "image_path": "11fd050c7836e50f9c8d71696e606b1b1ef5c767db8996739a597ba46ce164a5.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 219, + 374, + 391, + 390 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 506, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 133, + 408 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 395, + 181, + 407 + ], + "score": 0.92, + "content": "\\alpha _ { t } \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 394, + 338, + 408 + ], + "score": 1.0, + "content": "are constant hyper-parameters. When", + "type": "text" + }, + { + "bbox": [ + 339, + 397, + 350, + 406 + ], + "score": 0.86, + "content": "\\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "is small enough, we can approximate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 402, + 455, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 172, + 418 + ], + "score": 0.91, + "content": "x _ { T } ^ { 1 : N } \\sim \\mathcal { N } ( 0 , I )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 402, + 262, + 421 + ], + "score": 1.0, + "content": ". From here on we use", + "type": "text" + }, + { + "bbox": [ + 263, + 408, + 273, + 417 + ], + "score": 0.86, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 402, + 444, + 421 + ], + "score": 1.0, + "content": "to denote the full sequence at noising step", + "type": "text" + }, + { + "bbox": [ + 444, + 408, + 448, + 416 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 402, + 455, + 421 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 505, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 384, + 435 + ], + "score": 1.0, + "content": "In our context, conditioned motion synthesis models the distribution", + "type": "text" + }, + { + "bbox": [ + 385, + 423, + 416, + 435 + ], + "score": 0.92, + "content": "p ( x _ { 0 } | c )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "as the reversed diffu-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 248, + 446 + ], + "score": 1.0, + "content": "sion process of gradually cleaning", + "type": "text" + }, + { + "bbox": [ + 248, + 435, + 261, + 445 + ], + "score": 0.85, + "content": "x _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 434, + 353, + 446 + ], + "score": 1.0, + "content": ". Instead of predicting", + "type": "text" + }, + { + "bbox": [ + 353, + 435, + 362, + 445 + ], + "score": 0.84, + "content": "\\epsilon _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "as formulated by Ho et al. (2020),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "we follow Ramesh et al. (2022) and use an equivalent formulation to predict the signal itself, i.e.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 456, + 347, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 172, + 468 + ], + "score": 0.93, + "content": "\\hat { x } _ { 0 } = G ( x _ { t } , t , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 456, + 347, + 468 + ], + "score": 1.0, + "content": "with the simple objective (Ho et al., 2020),", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 473, + 406, + 488 + ], + "lines": [ + { + "bbox": [ + 205, + 473, + 406, + 488 + ], + "spans": [ + { + "bbox": [ + 205, + 473, + 406, + 488 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { s i m p l e } } = E _ { x _ { 0 } \\sim q ( x _ { 0 } | c ) , t \\sim [ 1 , T ] } [ \\| x _ { 0 } - G ( x _ { t } , t , c ) \\| _ { 2 } ^ { 2 } ]", + "type": "interline_equation", + "image_path": "3c8560f67f4f45f2c98fdecc22a60f73367236e6abdf30d227a76e25ffda0db5.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 205, + 473, + 406, + 488 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 494, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 509 + ], + "score": 1.0, + "content": "Geometric losses. In the motion domain, generative networks are standardly regularized using", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "geometric losses Petrovich et al. (2021); Shi et al. (2020). These losses enforce physical properties", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "and prevent artifacts, encouraging natural and coherent motion. In this work we experiment with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "three common geometric losses that regulate (1) positions (in case we predict rotations), (2) foot", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 538, + 213, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 213, + 550 + ], + "score": 1.0, + "content": "contact, and (3) velocities.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 543, + 386, + 577 + ], + "lines": [ + { + "bbox": [ + 224, + 543, + 386, + 577 + ], + "spans": [ + { + "bbox": [ + 224, + 543, + 386, + 577 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { \\mathrm { p o s } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\| F K ( x _ { 0 } ^ { i } ) - F K ( \\hat { x } _ { 0 } ^ { i } ) \\| _ { 2 } ^ { 2 } ,", + "type": "interline_equation", + "image_path": "83dc1a4ccd51b4d9ef24b1014f16302b9d7f9ed28ca5ce6413072eeed5daf8bd.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 224, + 543, + 386, + 560.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 224, + 560.0, + 386, + 577.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 586, + 412, + 620 + ], + "lines": [ + { + "bbox": [ + 198, + 586, + 412, + 620 + ], + "spans": [ + { + "bbox": [ + 198, + 586, + 412, + 620 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { \\mathrm { f o o t } } = \\frac { 1 } { N - 1 } \\sum _ { i = 1 } ^ { N - 1 } \\| ( F K ( \\hat { x } _ { 0 } ^ { i + 1 } ) - F K ( \\hat { x } _ { 0 } ^ { i } ) ) \\cdot f _ { i } \\| _ { 2 } ^ { 2 } ,", + "type": "interline_equation", + "image_path": "82159991daf78998072aa721251cf1cc292514d5cc6571d9271180114bcd0601.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 198, + 586, + 412, + 603.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 198, + 603.0, + 412, + 620.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 202, + 629, + 409, + 663 + ], + "lines": [ + { + "bbox": [ + 202, + 629, + 409, + 663 + ], + "spans": [ + { + "bbox": [ + 202, + 629, + 409, + 663 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { \\mathrm { v e l } } = \\frac { 1 } { N - 1 } \\sum _ { i = 1 } ^ { N - 1 } \\| ( x _ { 0 } ^ { i + 1 } - x _ { 0 } ^ { i } ) - ( \\hat { x } _ { 0 } ^ { i + 1 } - \\hat { x } _ { 0 } ^ { i } ) \\| _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "8b86f289003d23a8b1f4d9c86a75e25912e7dcd00b9e615165e4137137ac3796.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 629, + 409, + 646.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 202, + 646.0, + 409, + 663.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 246, + 678 + ], + "score": 1.0, + "content": "In case we predict joint rotations,", + "type": "text" + }, + { + "bbox": [ + 246, + 666, + 275, + 678 + ], + "score": 0.92, + "content": "F K ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "denotes the forward kinematic function converting joint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 400, + 690 + ], + "score": 1.0, + "content": "rotations into joint positions (otherwise, it denotes the identity function).", + "type": "text" + }, + { + "bbox": [ + 400, + 676, + 451, + 689 + ], + "score": 0.94, + "content": "f _ { i } \\in \\{ 0 , 1 \\} ^ { J }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "is the binary", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 239, + 700 + ], + "score": 1.0, + "content": "foot contact mask for each frame", + "type": "text" + }, + { + "bbox": [ + 239, + 689, + 243, + 698 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 687, + 504, + 700 + ], + "score": 1.0, + "content": ". Relevant only to feet, it indicates whether they touch the ground,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "and are set according to binary ground truth data (Shi et al., 2020). In essence, it mitigates the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "foot-sliding effect by nullifying velocities when touching the ground. Overall, our training loss is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 719, + 294, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 290, + 733 + ], + "score": 0.92, + "content": "{ \\mathcal { L } } = { \\mathcal { L } } _ { \\mathrm { s i m p l e } } + \\lambda _ { \\mathrm { p o s } } { \\bar { \\mathcal { L } } } _ { \\mathrm { p o s } } + { \\bar { \\lambda _ { \\mathrm { v e l } } } } { \\bar { \\mathcal { L } } } _ { \\mathrm { v e l } } + \\lambda _ { \\mathrm { f o o t } } { \\mathcal { L } } _ { \\mathrm { f o o t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 719, + 294, + 735 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 75, + 504, + 213 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 75, + 504, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 75, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 75, + 504, + 213 + ], + "score": 0.969, + "type": "image", + "image_path": "aceed64cf2b4fbe7a9a4e3115da6b56c1525a7c001a69455abf0263ac36383e4.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 75, + 504, + 121.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 121.0, + 504, + 167.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 167.0, + 504, + 213.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 220, + 505, + 298 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "Figure 2: (Left) Motion Diffusion Model (MDM) overview. The model is fed a motion sequence", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 226, + 509, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 127, + 244 + ], + "score": 0.91, + "content": "x _ { t } ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 226, + 165, + 248 + ], + "score": 1.0, + "content": "of length", + "type": "text" + }, + { + "bbox": [ + 166, + 232, + 176, + 242 + ], + "score": 0.8, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 226, + 243, + 248 + ], + "score": 1.0, + "content": "in a noising step", + "type": "text" + }, + { + "bbox": [ + 244, + 233, + 249, + 242 + ], + "score": 0.65, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 226, + 292, + 248 + ], + "score": 1.0, + "content": ", as well as", + "type": "text" + }, + { + "bbox": [ + 292, + 233, + 297, + 242 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 226, + 509, + 248 + ], + "score": 1.0, + "content": "itself and a conditioning code c. c, a CLIP (Radford", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 241, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 506, + 257 + ], + "score": 1.0, + "content": "et al., 2021) based textual embedding in this case, is first randomly masked for classifier-free learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 255, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 239, + 266 + ], + "score": 1.0, + "content": "and then projected together with", + "type": "text" + }, + { + "bbox": [ + 239, + 255, + 244, + 264 + ], + "score": 0.6, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 255, + 326, + 266 + ], + "score": 1.0, + "content": "into the input token", + "type": "text" + }, + { + "bbox": [ + 326, + 255, + 340, + 265 + ], + "score": 0.87, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 255, + 505, + 266 + ], + "score": 1.0, + "content": ". In each sampling step, the transformer-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 262, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 102, + 262, + 268, + 282 + ], + "score": 1.0, + "content": "encoder predicts the final clean motion", + "type": "text" + }, + { + "bbox": [ + 268, + 264, + 289, + 277 + ], + "score": 0.91, + "content": "\\hat { x } _ { 0 } ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 262, + 480, + 282 + ], + "score": 1.0, + "content": ". (Right) Sampling MDM. Given a condition", + "type": "text" + }, + { + "bbox": [ + 480, + 267, + 486, + 275 + ], + "score": 0.54, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 262, + 506, + 282 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 276, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 194, + 287 + ], + "score": 1.0, + "content": "sample random noise", + "type": "text" + }, + { + "bbox": [ + 194, + 277, + 207, + 287 + ], + "score": 0.86, + "content": "x _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 276, + 441, + 287 + ], + "score": 1.0, + "content": "at the dimensions of the desired motion, then iterate from", + "type": "text" + }, + { + "bbox": [ + 441, + 276, + 449, + 286 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 276, + 505, + 287 + ], + "score": 1.0, + "content": "to 1. At each", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 286, + 394, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 125, + 301 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 125, + 288, + 130, + 297 + ], + "score": 0.6, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 286, + 263, + 301 + ], + "score": 1.0, + "content": ", MDM predicts the clean sample", + "type": "text" + }, + { + "bbox": [ + 264, + 287, + 275, + 298 + ], + "score": 0.87, + "content": "{ \\hat { x } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 286, + 369, + 301 + ], + "score": 1.0, + "content": ", and diffuses it back to", + "type": "text" + }, + { + "bbox": [ + 370, + 289, + 389, + 298 + ], + "score": 0.89, + "content": "x _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 286, + 394, + 301 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 308, + 504, + 343 + ], + "lines": [], + "index": 11, + "bbox_fs": [ + 104, + 308, + 506, + 344 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 105, + 347, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 104, + 344, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 344, + 370, + 365 + ], + "score": 1.0, + "content": "Framework. Diffusion is modeled as a Markov noising process,", + "type": "text" + }, + { + "bbox": [ + 371, + 347, + 414, + 361 + ], + "score": 0.93, + "content": "\\{ x _ { t } ^ { 1 : N } \\} _ { t = 0 } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 344, + 446, + 365 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 446, + 348, + 466, + 361 + ], + "score": 0.92, + "content": "x _ { 0 } ^ { 1 : N }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 344, + 507, + 365 + ], + "score": 1.0, + "content": "is drawn", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 107, + 360, + 227, + 370 + ], + "spans": [ + { + "bbox": [ + 107, + 360, + 227, + 370 + ], + "score": 1.0, + "content": "from the data distribution and", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 344, + 507, + 370 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 374, + 391, + 390 + ], + "lines": [ + { + "bbox": [ + 219, + 374, + 391, + 390 + ], + "spans": [ + { + "bbox": [ + 219, + 374, + 391, + 390 + ], + "score": 0.93, + "content": "q ( x _ { t } ^ { 1 : N } | x _ { t - 1 } ^ { 1 : N } ) = \\mathcal { N } ( \\sqrt { \\alpha _ { t } } x _ { t - 1 } ^ { 1 : N } , ( 1 - \\alpha _ { t } ) I ) ,", + "type": "interline_equation", + "image_path": "11fd050c7836e50f9c8d71696e606b1b1ef5c767db8996739a597ba46ce164a5.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 219, + 374, + 391, + 390 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 506, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 133, + 408 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 395, + 181, + 407 + ], + "score": 0.92, + "content": "\\alpha _ { t } \\in ( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 394, + 338, + 408 + ], + "score": 1.0, + "content": "are constant hyper-parameters. When", + "type": "text" + }, + { + "bbox": [ + 339, + 397, + 350, + 406 + ], + "score": 0.86, + "content": "\\alpha _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "is small enough, we can approximate", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 402, + 455, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 172, + 418 + ], + "score": 0.91, + "content": "x _ { T } ^ { 1 : N } \\sim \\mathcal { N } ( 0 , I )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 402, + 262, + 421 + ], + "score": 1.0, + "content": ". From here on we use", + "type": "text" + }, + { + "bbox": [ + 263, + 408, + 273, + 417 + ], + "score": 0.86, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 402, + 444, + 421 + ], + "score": 1.0, + "content": "to denote the full sequence at noising step", + "type": "text" + }, + { + "bbox": [ + 444, + 408, + 448, + 416 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 402, + 455, + 421 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 394, + 505, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 505, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 384, + 435 + ], + "score": 1.0, + "content": "In our context, conditioned motion synthesis models the distribution", + "type": "text" + }, + { + "bbox": [ + 385, + 423, + 416, + 435 + ], + "score": 0.92, + "content": "p ( x _ { 0 } | c )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "as the reversed diffu-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 248, + 446 + ], + "score": 1.0, + "content": "sion process of gradually cleaning", + "type": "text" + }, + { + "bbox": [ + 248, + 435, + 261, + 445 + ], + "score": 0.85, + "content": "x _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 434, + 353, + 446 + ], + "score": 1.0, + "content": ". Instead of predicting", + "type": "text" + }, + { + "bbox": [ + 353, + 435, + 362, + 445 + ], + "score": 0.84, + "content": "\\epsilon _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "as formulated by Ho et al. (2020),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "we follow Ramesh et al. (2022) and use an equivalent formulation to predict the signal itself, i.e.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 456, + 347, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 172, + 468 + ], + "score": 0.93, + "content": "\\hat { x } _ { 0 } = G ( x _ { t } , t , c )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 456, + 347, + 468 + ], + "score": 1.0, + "content": "with the simple objective (Ho et al., 2020),", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 423, + 506, + 468 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 473, + 406, + 488 + ], + "lines": [ + { + "bbox": [ + 205, + 473, + 406, + 488 + ], + "spans": [ + { + "bbox": [ + 205, + 473, + 406, + 488 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { s i m p l e } } = E _ { x _ { 0 } \\sim q ( x _ { 0 } | c ) , t \\sim [ 1 , T ] } [ \\| x _ { 0 } - G ( x _ { t } , t , c ) \\| _ { 2 } ^ { 2 } ]", + "type": "interline_equation", + "image_path": "3c8560f67f4f45f2c98fdecc22a60f73367236e6abdf30d227a76e25ffda0db5.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 205, + 473, + 406, + 488 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 494, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 492, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 509 + ], + "score": 1.0, + "content": "Geometric losses. In the motion domain, generative networks are standardly regularized using", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "geometric losses Petrovich et al. (2021); Shi et al. (2020). These losses enforce physical properties", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "and prevent artifacts, encouraging natural and coherent motion. In this work we experiment with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "three common geometric losses that regulate (1) positions (in case we predict rotations), (2) foot", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 538, + 213, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 213, + 550 + ], + "score": 1.0, + "content": "contact, and (3) velocities.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 492, + 506, + 550 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 224, + 543, + 386, + 577 + ], + "lines": [ + { + "bbox": [ + 224, + 543, + 386, + 577 + ], + "spans": [ + { + "bbox": [ + 224, + 543, + 386, + 577 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { \\mathrm { p o s } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\| F K ( x _ { 0 } ^ { i } ) - F K ( \\hat { x } _ { 0 } ^ { i } ) \\| _ { 2 } ^ { 2 } ,", + "type": "interline_equation", + "image_path": "83dc1a4ccd51b4d9ef24b1014f16302b9d7f9ed28ca5ce6413072eeed5daf8bd.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 224, + 543, + 386, + 560.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 224, + 560.0, + 386, + 577.0 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 586, + 412, + 620 + ], + "lines": [ + { + "bbox": [ + 198, + 586, + 412, + 620 + ], + "spans": [ + { + "bbox": [ + 198, + 586, + 412, + 620 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { \\mathrm { f o o t } } = \\frac { 1 } { N - 1 } \\sum _ { i = 1 } ^ { N - 1 } \\| ( F K ( \\hat { x } _ { 0 } ^ { i + 1 } ) - F K ( \\hat { x } _ { 0 } ^ { i } ) ) \\cdot f _ { i } \\| _ { 2 } ^ { 2 } ,", + "type": "interline_equation", + "image_path": "82159991daf78998072aa721251cf1cc292514d5cc6571d9271180114bcd0601.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 198, + 586, + 412, + 603.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 198, + 603.0, + 412, + 620.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 202, + 629, + 409, + 663 + ], + "lines": [ + { + "bbox": [ + 202, + 629, + 409, + 663 + ], + "spans": [ + { + "bbox": [ + 202, + 629, + 409, + 663 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { \\mathrm { v e l } } = \\frac { 1 } { N - 1 } \\sum _ { i = 1 } ^ { N - 1 } \\| ( x _ { 0 } ^ { i + 1 } - x _ { 0 } ^ { i } ) - ( \\hat { x } _ { 0 } ^ { i + 1 } - \\hat { x } _ { 0 } ^ { i } ) \\| _ { 2 } ^ { 2 }", + "type": "interline_equation", + "image_path": "8b86f289003d23a8b1f4d9c86a75e25912e7dcd00b9e615165e4137137ac3796.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 202, + 629, + 409, + 646.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 202, + 646.0, + 409, + 663.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 246, + 678 + ], + "score": 1.0, + "content": "In case we predict joint rotations,", + "type": "text" + }, + { + "bbox": [ + 246, + 666, + 275, + 678 + ], + "score": 0.92, + "content": "F K ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "denotes the forward kinematic function converting joint", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 400, + 690 + ], + "score": 1.0, + "content": "rotations into joint positions (otherwise, it denotes the identity function).", + "type": "text" + }, + { + "bbox": [ + 400, + 676, + 451, + 689 + ], + "score": 0.94, + "content": "f _ { i } \\in \\{ 0 , 1 \\} ^ { J }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "is the binary", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 239, + 700 + ], + "score": 1.0, + "content": "foot contact mask for each frame", + "type": "text" + }, + { + "bbox": [ + 239, + 689, + 243, + 698 + ], + "score": 0.71, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 687, + 504, + 700 + ], + "score": 1.0, + "content": ". Relevant only to feet, it indicates whether they touch the ground,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "and are set according to binary ground truth data (Shi et al., 2020). In essence, it mitigates the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "foot-sliding effect by nullifying velocities when touching the ground. Overall, our training loss is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 719, + 294, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 290, + 733 + ], + "score": 0.92, + "content": "{ \\mathcal { L } } = { \\mathcal { L } } _ { \\mathrm { s i m p l e } } + \\lambda _ { \\mathrm { p o s } } { \\bar { \\mathcal { L } } } _ { \\mathrm { p o s } } + { \\bar { \\lambda _ { \\mathrm { v e l } } } } { \\bar { \\mathcal { L } } } _ { \\mathrm { v e l } } + \\lambda _ { \\mathrm { f o o t } } { \\mathcal { L } } _ { \\mathrm { f o o t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 719, + 294, + 735 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 665, + 506, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 77, + 490, + 245 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 77, + 490, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 77, + 490, + 245 + ], + "spans": [ + { + "bbox": [ + 108, + 77, + 490, + 245 + ], + "score": 0.976, + "type": "image", + "image_path": "3c301f5cc7f24dee220ca678e53e2e875ab4f7d26b316945108c82aaf3924b98.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 77, + 490, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 133.0, + 490, + 189.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 189.0, + 490, + 245.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 258, + 505, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "score": 1.0, + "content": "Figure 3: Editing applications. Light blue frames represent motion input and bronze frames are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "the generated motion. Motion in-betweening (left+center) can be performed conditioned on text or", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "without condition by the same model. Specific body part editing using text is demonstrated on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "right: the lower body joints are fixed to the input motion while the upper body is altered to fit the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 303, + 180, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 180, + 316 + ], + "score": 1.0, + "content": "input text prompt.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 371, + 327 + ], + "score": 1.0, + "content": "Model. Our model is illustrated in Figure 2. We implement", + "type": "text" + }, + { + "bbox": [ + 371, + 316, + 381, + 326 + ], + "score": 0.76, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "with a straightforward trans-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 326, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 504, + 339 + ], + "score": 1.0, + "content": "former (Vaswani et al., 2017) encoder-only architecture. The transformer architecture is temporally", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "score": 1.0, + "content": "aware, enabling learning and generating variable-length motions, and is well-proven for the motion", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 481, + 361 + ], + "score": 1.0, + "content": "domain (Petrovich et al., 2021; Duan et al., 2021; Aksan et al., 2021). The noise time-step", + "type": "text" + }, + { + "bbox": [ + 481, + 349, + 487, + 358 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 183, + 372 + ], + "score": 1.0, + "content": "the condition code", + "type": "text" + }, + { + "bbox": [ + 183, + 361, + 189, + 370 + ], + "score": 0.66, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "are each projected to the transformer dimension by separate feed-forward net-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 265, + 383 + ], + "score": 1.0, + "content": "works, then summed to yield the token", + "type": "text" + }, + { + "bbox": [ + 265, + 371, + 279, + 381 + ], + "score": 0.88, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 370, + 410, + 383 + ], + "score": 1.0, + "content": ". Each frame of the noised input", + "type": "text" + }, + { + "bbox": [ + 411, + 371, + 421, + 381 + ], + "score": 0.85, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "is linearly projected", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 381, + 504, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 439, + 394 + ], + "score": 1.0, + "content": "into the transformer dimension and summed with a standard positional embedding.", + "type": "text" + }, + { + "bbox": [ + 439, + 383, + 453, + 393 + ], + "score": 0.86, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 381, + 504, + 394 + ], + "score": 1.0, + "content": "and the pro-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 391, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 391, + 484, + 406 + ], + "score": 1.0, + "content": "jected frames are then fed to the encoder. Excluding the first output token (corresponding to", + "type": "text" + }, + { + "bbox": [ + 484, + 393, + 498, + 403 + ], + "score": 0.84, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 498, + 391, + 505, + 406 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "score": 1.0, + "content": "the encoder result is projected back to the original motion dimensions, and serves as the prediction", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 118, + 425 + ], + "score": 0.86, + "content": "\\scriptstyle { \\hat { x } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 414, + 366, + 428 + ], + "score": 1.0, + "content": ". We implement text-to-motion by encoding the text prompt to", + "type": "text" + }, + { + "bbox": [ + 366, + 416, + 372, + 424 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "with CLIP (Radford et al., 2021)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 426, + 388, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 388, + 437 + ], + "score": 1.0, + "content": "text encoder, and action-to-motion with learned embeddings per class.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 171, + 455 + ], + "score": 1.0, + "content": "Sampling from", + "type": "text" + }, + { + "bbox": [ + 172, + 442, + 203, + 454 + ], + "score": 0.93, + "content": "p ( x _ { 0 } | c )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "is done in an iterative manner, according to Ho et al. (2020). In every time", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 125, + 466 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 125, + 454, + 130, + 463 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 453, + 244, + 466 + ], + "score": 1.0, + "content": "we predict the clean sample", + "type": "text" + }, + { + "bbox": [ + 245, + 453, + 311, + 465 + ], + "score": 0.92, + "content": "\\hat { \\boldsymbol { x } } _ { 0 } = \\boldsymbol { G } ( \\boldsymbol { x } _ { t } , t , \\boldsymbol { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 453, + 392, + 466 + ], + "score": 1.0, + "content": "and noise it back to", + "type": "text" + }, + { + "bbox": [ + 392, + 455, + 413, + 464 + ], + "score": 0.88, + "content": "x _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 453, + 505, + 466 + ], + "score": 1.0, + "content": ". This is repeated from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 464, + 504, + 476 + ], + "spans": [ + { + "bbox": [ + 107, + 464, + 132, + 474 + ], + "score": 0.9, + "content": "t = T", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 464, + 153, + 476 + ], + "score": 1.0, + "content": "until", + "type": "text" + }, + { + "bbox": [ + 154, + 465, + 165, + 475 + ], + "score": 0.85, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 464, + 358, + 476 + ], + "score": 1.0, + "content": "is achieved (Figure 2 right). We train our model", + "type": "text" + }, + { + "bbox": [ + 358, + 464, + 367, + 474 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 464, + 504, + 476 + ], + "score": 1.0, + "content": "using classifier-free guidance (Ho", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 475, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 237, + 487 + ], + "score": 1.0, + "content": "& Salimans, 2022). In practice,", + "type": "text" + }, + { + "bbox": [ + 237, + 475, + 247, + 485 + ], + "score": 0.79, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 475, + 504, + 487 + ], + "score": 1.0, + "content": "learns both the conditioned and the unconditioned distributions", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 189, + 499 + ], + "score": 1.0, + "content": "by randomly setting", + "type": "text" + }, + { + "bbox": [ + 189, + 486, + 213, + 496 + ], + "score": 0.91, + "content": "c = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 484, + 228, + 499 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 228, + 486, + 248, + 496 + ], + "score": 0.89, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 484, + 350, + 499 + ], + "score": 1.0, + "content": "of the samples, such that", + "type": "text" + }, + { + "bbox": [ + 350, + 486, + 393, + 497 + ], + "score": 0.93, + "content": "G ( x _ { t } , t , \\emptyset )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 484, + 451, + 499 + ], + "score": 1.0, + "content": "approximates", + "type": "text" + }, + { + "bbox": [ + 451, + 486, + 475, + 498 + ], + "score": 0.92, + "content": "p ( x _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 484, + 505, + 499 + ], + "score": 1.0, + "content": ". Then,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 170, + 509 + ], + "score": 1.0, + "content": "when sampling", + "type": "text" + }, + { + "bbox": [ + 171, + 497, + 180, + 507 + ], + "score": 0.77, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "we can trade-off diversity and fidelity by interpolating or even extrapolating the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 507, + 426, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 182, + 521 + ], + "score": 1.0, + "content": "two variants using", + "type": "text" + }, + { + "bbox": [ + 182, + 507, + 422, + 520 + ], + "score": 0.86, + "content": "\\boldsymbol { s } \\colon G _ { s } ( x _ { t } , t , c ) = G ( x _ { t } , \\dot { t , } \\emptyset ) + s \\cdot ( \\dot { G } ( x _ { t } , t , c ) \\bar { - } G ( x _ { t } , t , \\emptyset ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 507, + 426, + 521 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 523, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 538 + ], + "score": 1.0, + "content": "Editing. We enable motion in-betweening in the temporal domain, and body part editing in the spa-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "tial domain, by adapting diffusion inpainting to motion data. Editing is done only during sampling,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "without any training involved. Given a subset of the motion sequence inputs, when sampling the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 328, + 569 + ], + "score": 1.0, + "content": "model (Figure 2 right), at each iteration we overwrite", + "type": "text" + }, + { + "bbox": [ + 329, + 558, + 340, + 569 + ], + "score": 0.88, + "content": "\\scriptstyle { \\hat { x } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "with the input part of the motion. This", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "encourages the generation to remain coherent to original input, while completing the missing parts.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "In the temporal setting, the prefix and suffix frames of the motion sequence are the input, and we", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 591, + 504, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 504, + 603 + ], + "score": 1.0, + "content": "solve a motion in-betweening problem (Harvey et al., 2020). Editing can be done either condition-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 600, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 248, + 615 + ], + "score": 1.0, + "content": "ally or unconditionally (by setting", + "type": "text" + }, + { + "bbox": [ + 248, + 601, + 275, + 612 + ], + "score": 0.89, + "content": "c = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 600, + 505, + 615 + ], + "score": 1.0, + "content": "). In the spatial setting, we show that body parts can be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 266, + 626 + ], + "score": 1.0, + "content": "re-synthesized according to a condition", + "type": "text" + }, + { + "bbox": [ + 267, + 615, + 272, + 622 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "while keeping the rest intact, through the use of the same", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 623, + 198, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 198, + 636 + ], + "score": 1.0, + "content": "completion technique.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 200, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 201, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 201, + 660 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "We implement MDM for three motion generation tasks: Text-to-Motion(4.1), Action-to-Motion(4.2)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "and unconditioned generation(5.2. Each sub-section reviews the data and metrics of the used bench-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "marks, provides implementation details, and presents qualitative and quantitative results. Then, we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "show implementations of motion in-betweening (both conditioned and unconditioned) and body-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "part editing by adapting diffusion inpainting to motion (5.1). Our models have been trained with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 149, + 731 + ], + "score": 0.89, + "content": "T = 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "noising steps and a cosine noise schedule. In Appendix A.1, we experiment with differ-", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 77, + 490, + 245 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 77, + 490, + 245 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 77, + 490, + 245 + ], + "spans": [ + { + "bbox": [ + 108, + 77, + 490, + 245 + ], + "score": 0.976, + "type": "image", + "image_path": "3c301f5cc7f24dee220ca678e53e2e875ab4f7d26b316945108c82aaf3924b98.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 77, + 490, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 133.0, + 490, + 189.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 189.0, + 490, + 245.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 258, + 505, + 313 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 506, + 272 + ], + "score": 1.0, + "content": "Figure 3: Editing applications. Light blue frames represent motion input and bronze frames are", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "the generated motion. Motion in-betweening (left+center) can be performed conditioned on text or", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "without condition by the same model. Specific body part editing using text is demonstrated on the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "right: the lower body joints are fixed to the input motion while the upper body is altered to fit the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 303, + 180, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 180, + 316 + ], + "score": 1.0, + "content": "input text prompt.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + } + ], + "index": 3.0 + }, + { + "type": "text", + "bbox": [ + 106, + 315, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 371, + 327 + ], + "score": 1.0, + "content": "Model. Our model is illustrated in Figure 2. We implement", + "type": "text" + }, + { + "bbox": [ + 371, + 316, + 381, + 326 + ], + "score": 0.76, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "with a straightforward trans-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 326, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 504, + 339 + ], + "score": 1.0, + "content": "former (Vaswani et al., 2017) encoder-only architecture. The transformer architecture is temporally", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 349 + ], + "score": 1.0, + "content": "aware, enabling learning and generating variable-length motions, and is well-proven for the motion", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 481, + 361 + ], + "score": 1.0, + "content": "domain (Petrovich et al., 2021; Duan et al., 2021; Aksan et al., 2021). The noise time-step", + "type": "text" + }, + { + "bbox": [ + 481, + 349, + 487, + 358 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 183, + 372 + ], + "score": 1.0, + "content": "the condition code", + "type": "text" + }, + { + "bbox": [ + 183, + 361, + 189, + 370 + ], + "score": 0.66, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "are each projected to the transformer dimension by separate feed-forward net-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 265, + 383 + ], + "score": 1.0, + "content": "works, then summed to yield the token", + "type": "text" + }, + { + "bbox": [ + 265, + 371, + 279, + 381 + ], + "score": 0.88, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 370, + 410, + 383 + ], + "score": 1.0, + "content": ". Each frame of the noised input", + "type": "text" + }, + { + "bbox": [ + 411, + 371, + 421, + 381 + ], + "score": 0.85, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "is linearly projected", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 381, + 504, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 439, + 394 + ], + "score": 1.0, + "content": "into the transformer dimension and summed with a standard positional embedding.", + "type": "text" + }, + { + "bbox": [ + 439, + 383, + 453, + 393 + ], + "score": 0.86, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 381, + 504, + 394 + ], + "score": 1.0, + "content": "and the pro-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 391, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 391, + 484, + 406 + ], + "score": 1.0, + "content": "jected frames are then fed to the encoder. Excluding the first output token (corresponding to", + "type": "text" + }, + { + "bbox": [ + 484, + 393, + 498, + 403 + ], + "score": 0.84, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 498, + 391, + 505, + 406 + ], + "score": 1.0, + "content": "),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "score": 1.0, + "content": "the encoder result is projected back to the original motion dimensions, and serves as the prediction", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 118, + 425 + ], + "score": 0.86, + "content": "\\scriptstyle { \\hat { x } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 414, + 366, + 428 + ], + "score": 1.0, + "content": ". We implement text-to-motion by encoding the text prompt to", + "type": "text" + }, + { + "bbox": [ + 366, + 416, + 372, + 424 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 414, + 506, + 428 + ], + "score": 1.0, + "content": "with CLIP (Radford et al., 2021)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 426, + 388, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 388, + 437 + ], + "score": 1.0, + "content": "text encoder, and action-to-motion with learned embeddings per class.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13, + "bbox_fs": [ + 104, + 315, + 506, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 171, + 455 + ], + "score": 1.0, + "content": "Sampling from", + "type": "text" + }, + { + "bbox": [ + 172, + 442, + 203, + 454 + ], + "score": 0.93, + "content": "p ( x _ { 0 } | c )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 441, + 505, + 455 + ], + "score": 1.0, + "content": "is done in an iterative manner, according to Ho et al. (2020). In every time", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 125, + 466 + ], + "score": 1.0, + "content": "step", + "type": "text" + }, + { + "bbox": [ + 125, + 454, + 130, + 463 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 453, + 244, + 466 + ], + "score": 1.0, + "content": "we predict the clean sample", + "type": "text" + }, + { + "bbox": [ + 245, + 453, + 311, + 465 + ], + "score": 0.92, + "content": "\\hat { \\boldsymbol { x } } _ { 0 } = \\boldsymbol { G } ( \\boldsymbol { x } _ { t } , t , \\boldsymbol { c } )", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 453, + 392, + 466 + ], + "score": 1.0, + "content": "and noise it back to", + "type": "text" + }, + { + "bbox": [ + 392, + 455, + 413, + 464 + ], + "score": 0.88, + "content": "x _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 453, + 505, + 466 + ], + "score": 1.0, + "content": ". This is repeated from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 107, + 464, + 504, + 476 + ], + "spans": [ + { + "bbox": [ + 107, + 464, + 132, + 474 + ], + "score": 0.9, + "content": "t = T", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 464, + 153, + 476 + ], + "score": 1.0, + "content": "until", + "type": "text" + }, + { + "bbox": [ + 154, + 465, + 165, + 475 + ], + "score": 0.85, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 464, + 358, + 476 + ], + "score": 1.0, + "content": "is achieved (Figure 2 right). We train our model", + "type": "text" + }, + { + "bbox": [ + 358, + 464, + 367, + 474 + ], + "score": 0.82, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 464, + 504, + 476 + ], + "score": 1.0, + "content": "using classifier-free guidance (Ho", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 475, + 504, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 237, + 487 + ], + "score": 1.0, + "content": "& Salimans, 2022). In practice,", + "type": "text" + }, + { + "bbox": [ + 237, + 475, + 247, + 485 + ], + "score": 0.79, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 475, + 504, + 487 + ], + "score": 1.0, + "content": "learns both the conditioned and the unconditioned distributions", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 189, + 499 + ], + "score": 1.0, + "content": "by randomly setting", + "type": "text" + }, + { + "bbox": [ + 189, + 486, + 213, + 496 + ], + "score": 0.91, + "content": "c = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 484, + 228, + 499 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 228, + 486, + 248, + 496 + ], + "score": 0.89, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 484, + 350, + 499 + ], + "score": 1.0, + "content": "of the samples, such that", + "type": "text" + }, + { + "bbox": [ + 350, + 486, + 393, + 497 + ], + "score": 0.93, + "content": "G ( x _ { t } , t , \\emptyset )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 484, + 451, + 499 + ], + "score": 1.0, + "content": "approximates", + "type": "text" + }, + { + "bbox": [ + 451, + 486, + 475, + 498 + ], + "score": 0.92, + "content": "p ( x _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 484, + 505, + 499 + ], + "score": 1.0, + "content": ". Then,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 170, + 509 + ], + "score": 1.0, + "content": "when sampling", + "type": "text" + }, + { + "bbox": [ + 171, + 497, + 180, + 507 + ], + "score": 0.77, + "content": "G", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "we can trade-off diversity and fidelity by interpolating or even extrapolating the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 507, + 426, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 182, + 521 + ], + "score": 1.0, + "content": "two variants using", + "type": "text" + }, + { + "bbox": [ + 182, + 507, + 422, + 520 + ], + "score": 0.86, + "content": "\\boldsymbol { s } \\colon G _ { s } ( x _ { t } , t , c ) = G ( x _ { t } , \\dot { t , } \\emptyset ) + s \\cdot ( \\dot { G } ( x _ { t } , t , c ) \\bar { - } G ( x _ { t } , t , \\emptyset ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 507, + 426, + 521 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 441, + 505, + 521 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 523, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 538 + ], + "score": 1.0, + "content": "Editing. We enable motion in-betweening in the temporal domain, and body part editing in the spa-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "tial domain, by adapting diffusion inpainting to motion data. Editing is done only during sampling,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "without any training involved. Given a subset of the motion sequence inputs, when sampling the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 328, + 569 + ], + "score": 1.0, + "content": "model (Figure 2 right), at each iteration we overwrite", + "type": "text" + }, + { + "bbox": [ + 329, + 558, + 340, + 569 + ], + "score": 0.88, + "content": "\\scriptstyle { \\hat { x } } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "with the input part of the motion. This", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "encourages the generation to remain coherent to original input, while completing the missing parts.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "In the temporal setting, the prefix and suffix frames of the motion sequence are the input, and we", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 591, + 504, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 504, + 603 + ], + "score": 1.0, + "content": "solve a motion in-betweening problem (Harvey et al., 2020). Editing can be done either condition-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 600, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 248, + 615 + ], + "score": 1.0, + "content": "ally or unconditionally (by setting", + "type": "text" + }, + { + "bbox": [ + 248, + 601, + 275, + 612 + ], + "score": 0.89, + "content": "c = \\emptyset", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 600, + 505, + 615 + ], + "score": 1.0, + "content": "). In the spatial setting, we show that body parts can be", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 266, + 626 + ], + "score": 1.0, + "content": "re-synthesized according to a condition", + "type": "text" + }, + { + "bbox": [ + 267, + 615, + 272, + 622 + ], + "score": 0.7, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "while keeping the rest intact, through the use of the same", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 623, + 198, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 198, + 636 + ], + "score": 1.0, + "content": "completion technique.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 523, + 505, + 636 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 646, + 200, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 201, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 201, + 660 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "We implement MDM for three motion generation tasks: Text-to-Motion(4.1), Action-to-Motion(4.2)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "and unconditioned generation(5.2. Each sub-section reviews the data and metrics of the used bench-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "marks, provides implementation details, and presents qualitative and quantitative results. Then, we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "show implementations of motion in-betweening (both conditioned and unconditioned) and body-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "part editing by adapting diffusion inpainting to motion (5.1). Our models have been trained with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 149, + 731 + ], + "score": 0.89, + "content": "T = 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "noising steps and a cosine noise schedule. In Appendix A.1, we experiment with differ-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 159, + 95 + ], + "score": 1.0, + "content": "ent values of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 159, + 83, + 168, + 92 + ], + "score": 0.73, + "content": "T", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 168, + 82, + 506, + 95 + ], + "score": 1.0, + "content": ". All of them have been trained on a single NVIDIA GeForce RTX 2080 Ti GPU for a", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 200, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 200, + 106 + ], + "score": 1.0, + "content": "period of about 3 days.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 666, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 159, + 95 + ], + "score": 1.0, + "content": "ent values of", + "type": "text" + }, + { + "bbox": [ + 159, + 83, + 168, + 92 + ], + "score": 0.73, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 82, + 506, + 95 + ], + "score": 1.0, + "content": ". All of them have been trained on a single NVIDIA GeForce RTX 2080 Ti GPU for a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 200, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 200, + 106 + ], + "score": 1.0, + "content": "period of about 3 days.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 115, + 210, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 114, + 211, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 211, + 128 + ], + "score": 1.0, + "content": "4.1 TEXT-TO-MOTION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 131, + 505, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "Text-to-motion is the task of generating motion given an input text prompt. The output motion is ex-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "pected to be both implementing the textual description, and a valid sample from the data distribution", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "(i.e. adhering to general human abilities and the rules of physics). In addition, for each text prompt,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "we also expect a distribution of motions matching it, rather than just a single result. We evaluate our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 175, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 188 + ], + "score": 1.0, + "content": "model using two leading benchmarks - KIT (Plappert et al., 2016) and HumanML3D (Guo et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "2022a), over the set of metrics suggested by Guo et al. (2022a): R-precision and Multimodal-Dist", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 104, + 197, + 386, + 210 + ], + "score": 1.0, + "content": "measure the relevancy of the generated motions to the input prompts,", + "type": "text" + }, + { + "bbox": [ + 386, + 198, + 405, + 208 + ], + "score": 0.3, + "content": "F I D", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "measures the dissimilar-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "score": 1.0, + "content": "ity between the generated and ground truth distributions (in latent space), Diversity measures the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "variability in the resulting motion distribution, and MultiModality is the average variance given a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "single text prompt. For the full implementation of the metrics, please refer to Guo et al. (2022a).", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "We use HumanML3D as a platform to compare different backbones of our model, discovering that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 265 + ], + "score": 1.0, + "content": "the diffusion framework is relatively agnostic to this attribute. In addition, we conduct a user study", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 263, + 355, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 355, + 276 + ], + "score": 1.0, + "content": "comparing our model to current art and ground truth motions.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "Data. HumanML3D is a recent dataset, textually re-annotating motion capture from the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "AMASS (Mahmood et al., 2019) and HumanAct12 (Guo et al., 2020) collections. It contains 14, 616", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "score": 1.0, + "content": "motions annotated by 44, 970 textual descriptions. In addition, it suggests a redundant data repre-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "sentation including a concatenation of root velocity, joint positions, joint velocities, joint rotations", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "and the foot contact binary labels. We also use in this section the same representation for the KIT", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "dataset, brought by the same publishers. Although limited in the number (3, 911) and the diversity", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "of samples, most of the text-to-motion research is based on KIT, hence we view it as important to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 357, + 207, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 207, + 369 + ], + "score": 1.0, + "content": "evaluate using it as well.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "Implementation. In addition to our Transformer encoder-only backbone (Section 3), we experiment", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 379, + 397 + ], + "score": 1.0, + "content": "MDM with three more backbones: (1) Transformer decoder injects", + "type": "text" + }, + { + "bbox": [ + 380, + 386, + 394, + 396 + ], + "score": 0.88, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "through the cross-attention", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 343, + 408 + ], + "score": 1.0, + "content": "layer, instead of as an input token. (2) Transformer decoder", + "type": "text" + }, + { + "bbox": [ + 343, + 397, + 352, + 406 + ], + "score": 0.7, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 396, + 427, + 408 + ], + "score": 1.0, + "content": "input token, where", + "type": "text" + }, + { + "bbox": [ + 428, + 397, + 441, + 407 + ], + "score": 0.87, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "is injected both", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 294, + 419 + ], + "score": 1.0, + "content": "ways, (3) GRU (Cho et al., 2014) concatenate", + "type": "text" + }, + { + "bbox": [ + 294, + 408, + 308, + 418 + ], + "score": 0.87, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "to each input frame , and (4) a U-net adaptation", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "for motion data (Table 1). U-net is adapted to motion by replacing the 2D convolution filters with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "1D convolutions in the temporal axis, such that joints are considered as channels. This is due to the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 438, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 453 + ], + "score": 1.0, + "content": "irregular behavior of the joint axis. Our models were trained with batch size 64, 8 layers (except", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "GRU that was optimal at 2), and latent dimension 512. To encode the text we use a frozen CLIP-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "ViT-B/32 model. In addition, we experiment with replacing CLIP with sentence-BERT (Reimers &", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 473, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 485 + ], + "score": 1.0, + "content": "Gurevych, 2019) - a BERT-based (Devlin et al., 2019) text encoder. The full details can be found", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 394, + 496 + ], + "score": 1.0, + "content": "in our published code1 and Appendix C. Each model was trained for", + "type": "text" + }, + { + "bbox": [ + 394, + 483, + 420, + 494 + ], + "score": 0.84, + "content": "5 0 0 K", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "steps, after which a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "checkpoint was chosen that minimizes the FID metric to be reported. Since foot contact and joint", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "locations are explicitly represented in HumanML3D, we don’t apply geometric losses in this section.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 287, + 529 + ], + "score": 1.0, + "content": "We evaluate our models with guidance-scale", + "type": "text" + }, + { + "bbox": [ + 287, + 516, + 320, + 527 + ], + "score": 0.88, + "content": "s = 2 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "which provides a diversity-fidelity sweet spot", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 527, + 152, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 152, + 540 + ], + "score": 1.0, + "content": "(Figure 4).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "Quantitative evaluation. We evaluate and compare our models to current art (JL2P Ahuja &", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "Morency (2019), Text2Gesture (Bhattacharya et al., 2021), and T2M (Guo et al., 2022a)) with the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "metrics suggested by Guo et al. (2022a). As can be seen, MDM achieves state-of-the-art results in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "FID, Diversity, and MultiModality, indicating high diversity per input text prompt, and high-quality", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 588, + 322, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 322, + 601 + ], + "score": 1.0, + "content": "samples, as can also be seen qualitatively in Figure 1.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "User study. We asked 31 users to choose between MDM and state-of-the-art works in a side-by-side", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "view, with both samples generated from the same text prompt randomly sampled from the KIT test", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "set. We repeated this process with 10 samples per model and 10 repetitions per sample. This user", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "study enabled a comparison with the recent TEMOS model (Petrovich et al., 2022), which was not", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "included in the HumanML3D benchmark. Fig. 4 shows that most of the time, MDM was preferred", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 659, + 490, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 410, + 672 + ], + "score": 1.0, + "content": "over the compared models, and even preferred over ground truth samples in", + "type": "text" + }, + { + "bbox": [ + 410, + 659, + 438, + 670 + ], + "score": 0.88, + "content": "4 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 659, + 490, + 672 + ], + "score": 1.0, + "content": "of the cases.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5 + }, + { + "type": "title", + "bbox": [ + 108, + 672, + 221, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 223, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 223, + 684 + ], + "score": 1.0, + "content": "4.2 ACTION-TO-MOTION", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Action-to-motion is the task of generating motion given an input action class, represented by a scalar.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "The output motion should faithfully animate the input action, and at the same time be natural and", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "reflect the distribution of the dataset on which the model is trained. Two dataset are commonly used", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 720, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 504, + 732 + ], + "score": 1.0, + "content": "to evaluate action-to-motion models: HumanAct12 (Guo et al., 2020) and UESTC (Ji et al., 2018).", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 506, + 106 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 115, + 210, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 114, + 211, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 211, + 128 + ], + "score": 1.0, + "content": "4.1 TEXT-TO-MOTION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 131, + 505, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "Text-to-motion is the task of generating motion given an input text prompt. The output motion is ex-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "pected to be both implementing the textual description, and a valid sample from the data distribution", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "(i.e. adhering to general human abilities and the rules of physics). In addition, for each text prompt,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "we also expect a distribution of motions matching it, rather than just a single result. We evaluate our", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 175, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 188 + ], + "score": 1.0, + "content": "model using two leading benchmarks - KIT (Plappert et al., 2016) and HumanML3D (Guo et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 506, + 199 + ], + "score": 1.0, + "content": "2022a), over the set of metrics suggested by Guo et al. (2022a): R-precision and Multimodal-Dist", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 197, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 104, + 197, + 386, + 210 + ], + "score": 1.0, + "content": "measure the relevancy of the generated motions to the input prompts,", + "type": "text" + }, + { + "bbox": [ + 386, + 198, + 405, + 208 + ], + "score": 0.3, + "content": "F I D", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 197, + 505, + 210 + ], + "score": 1.0, + "content": "measures the dissimilar-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 221 + ], + "score": 1.0, + "content": "ity between the generated and ground truth distributions (in latent space), Diversity measures the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "variability in the resulting motion distribution, and MultiModality is the average variance given a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "single text prompt. For the full implementation of the metrics, please refer to Guo et al. (2022a).", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 506, + 254 + ], + "score": 1.0, + "content": "We use HumanML3D as a platform to compare different backbones of our model, discovering that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 505, + 265 + ], + "score": 1.0, + "content": "the diffusion framework is relatively agnostic to this attribute. In addition, we conduct a user study", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 263, + 355, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 355, + 276 + ], + "score": 1.0, + "content": "comparing our model to current art and ground truth motions.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9, + "bbox_fs": [ + 104, + 131, + 506, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 280, + 505, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "Data. HumanML3D is a recent dataset, textually re-annotating motion capture from the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "AMASS (Mahmood et al., 2019) and HumanAct12 (Guo et al., 2020) collections. It contains 14, 616", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "score": 1.0, + "content": "motions annotated by 44, 970 textual descriptions. In addition, it suggests a redundant data repre-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "sentation including a concatenation of root velocity, joint positions, joint velocities, joint rotations", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "and the foot contact binary labels. We also use in this section the same representation for the KIT", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "dataset, brought by the same publishers. Although limited in the number (3, 911) and the diversity", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "of samples, most of the text-to-motion research is based on KIT, hence we view it as important to", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 357, + 207, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 207, + 369 + ], + "score": 1.0, + "content": "evaluate using it as well.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 280, + 505, + 369 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 505, + 386 + ], + "score": 1.0, + "content": "Implementation. In addition to our Transformer encoder-only backbone (Section 3), we experiment", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 379, + 397 + ], + "score": 1.0, + "content": "MDM with three more backbones: (1) Transformer decoder injects", + "type": "text" + }, + { + "bbox": [ + 380, + 386, + 394, + 396 + ], + "score": 0.88, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "through the cross-attention", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 343, + 408 + ], + "score": 1.0, + "content": "layer, instead of as an input token. (2) Transformer decoder", + "type": "text" + }, + { + "bbox": [ + 343, + 397, + 352, + 406 + ], + "score": 0.7, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 396, + 427, + 408 + ], + "score": 1.0, + "content": "input token, where", + "type": "text" + }, + { + "bbox": [ + 428, + 397, + 441, + 407 + ], + "score": 0.87, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "is injected both", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 294, + 419 + ], + "score": 1.0, + "content": "ways, (3) GRU (Cho et al., 2014) concatenate", + "type": "text" + }, + { + "bbox": [ + 294, + 408, + 308, + 418 + ], + "score": 0.87, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "to each input frame , and (4) a U-net adaptation", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 431 + ], + "score": 1.0, + "content": "for motion data (Table 1). U-net is adapted to motion by replacing the 2D convolution filters with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 427, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 506, + 442 + ], + "score": 1.0, + "content": "1D convolutions in the temporal axis, such that joints are considered as channels. This is due to the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 438, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 453 + ], + "score": 1.0, + "content": "irregular behavior of the joint axis. Our models were trained with batch size 64, 8 layers (except", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "GRU that was optimal at 2), and latent dimension 512. To encode the text we use a frozen CLIP-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "ViT-B/32 model. In addition, we experiment with replacing CLIP with sentence-BERT (Reimers &", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 473, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 485 + ], + "score": 1.0, + "content": "Gurevych, 2019) - a BERT-based (Devlin et al., 2019) text encoder. The full details can be found", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 394, + 496 + ], + "score": 1.0, + "content": "in our published code1 and Appendix C. Each model was trained for", + "type": "text" + }, + { + "bbox": [ + 394, + 483, + 420, + 494 + ], + "score": 0.84, + "content": "5 0 0 K", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "steps, after which a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 505, + 506 + ], + "score": 1.0, + "content": "checkpoint was chosen that minimizes the FID metric to be reported. Since foot contact and joint", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "locations are explicitly represented in HumanML3D, we don’t apply geometric losses in this section.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 287, + 529 + ], + "score": 1.0, + "content": "We evaluate our models with guidance-scale", + "type": "text" + }, + { + "bbox": [ + 287, + 516, + 320, + 527 + ], + "score": 0.88, + "content": "s = 2 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "which provides a diversity-fidelity sweet spot", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 527, + 152, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 152, + 540 + ], + "score": 1.0, + "content": "(Figure 4).", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 373, + 506, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "Quantitative evaluation. We evaluate and compare our models to current art (JL2P Ahuja &", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "Morency (2019), Text2Gesture (Bhattacharya et al., 2021), and T2M (Guo et al., 2022a)) with the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 579 + ], + "score": 1.0, + "content": "metrics suggested by Guo et al. (2022a). As can be seen, MDM achieves state-of-the-art results in", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "FID, Diversity, and MultiModality, indicating high diversity per input text prompt, and high-quality", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 588, + 322, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 322, + 601 + ], + "score": 1.0, + "content": "samples, as can also be seen qualitatively in Figure 1.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 544, + 506, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "User study. We asked 31 users to choose between MDM and state-of-the-art works in a side-by-side", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "view, with both samples generated from the same text prompt randomly sampled from the KIT test", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "set. We repeated this process with 10 samples per model and 10 repetitions per sample. This user", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "study enabled a comparison with the recent TEMOS model (Petrovich et al., 2022), which was not", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 660 + ], + "score": 1.0, + "content": "included in the HumanML3D benchmark. Fig. 4 shows that most of the time, MDM was preferred", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 659, + 490, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 410, + 672 + ], + "score": 1.0, + "content": "over the compared models, and even preferred over ground truth samples in", + "type": "text" + }, + { + "bbox": [ + 410, + 659, + 438, + 670 + ], + "score": 0.88, + "content": "4 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 659, + 490, + 672 + ], + "score": 1.0, + "content": "of the cases.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 604, + 506, + 672 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 672, + 221, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 223, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 223, + 684 + ], + "score": 1.0, + "content": "4.2 ACTION-TO-MOTION", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "Action-to-motion is the task of generating motion given an input action class, represented by a scalar.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "The output motion should faithfully animate the input action, and at the same time be natural and", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "reflect the distribution of the dataset on which the model is trained. Two dataset are commonly used", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 720, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 504, + 732 + ], + "score": 1.0, + "content": "to evaluate action-to-motion models: HumanAct12 (Guo et al., 2020) and UESTC (Ji et al., 2018).", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 501, + 215 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 70, + 501, + 215 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 70, + 501, + 215 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 501, + 215 + ], + "score": 0.984, + "html": "
MethodRPrecision (top 3)↑FID↓Multimodal Dist↓Diversity→Multimodality↑
Real0.797±.0020.002±.0002.974±.0089.503±.065
JL2P0.486±.00211.02±.0465.296±.0087.676±.0581
Text2Gesture0.345±.0027.664±.0306.030±.0086.409±.0711
T2M0.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM (ours)0.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
+ sent-BERT0.609±.0060.586±.0365.504±.039.666±.0952.707±.188
MDM (decoder)0.608±.0050.767±.0855.507±.0209.176±.0702.927±.125
+ input token0.621±.0050.567±.0515.424±.0229.425±.0602.834±.095
MDM (U-net)0.603±.0061.137±.0085.629±.0328.958±.0982.636±.214
MDM (GRU)0.645±.0054.569±.1505.325±.0267.688±.0821.2646±.024
", + "type": "table", + "image_path": "a8390e456fc75cdd5ea1504b7990139329e00b053fc9fa17c46ccd72412e5cff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 70, + 501, + 118.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 118.33333333333334, + 501, + 166.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 166.66666666666669, + 501, + 215.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 223, + 505, + 268 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "Table 1: Quantitative results on the HumanML3D test set. All methods use the real motion length", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 201, + 246 + ], + "score": 1.0, + "content": "from the ground truth.", + "type": "text" + }, + { + "bbox": [ + 201, + 235, + 220, + 245 + ], + "score": 0.82, + "content": "\\cdot _ { } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "means results are better if the metric is closer to the real distribution.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 245, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 418, + 258 + ], + "score": 1.0, + "content": "We run all the evaluation 20 times (except MultiModality runs 5 times) and", + "type": "text" + }, + { + "bbox": [ + 418, + 246, + 429, + 256 + ], + "score": 0.67, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 245, + 484, + 258 + ], + "score": 1.0, + "content": "indicates the", + "type": "text" + }, + { + "bbox": [ + 484, + 245, + 504, + 256 + ], + "score": 0.85, + "content": "9 5 \\%", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 257, + 293, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 293, + 268 + ], + "score": 1.0, + "content": "confidence interval. Bold indicates best result.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 108, + 280, + 501, + 376 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 280, + 501, + 376 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 280, + 501, + 376 + ], + "spans": [ + { + "bbox": [ + 108, + 280, + 501, + 376 + ], + "score": 0.983, + "html": "
MethodR Precision (top 3)↑FID↓Multimodal Dist↓Diversity→Multimodality↑
Real0.779±.0060.031±.0042.788±.01211.08±.097-
JL2P0.483±.0056.545±.0725.147±.0309.073±.100-
Text2Gesture0.338±.00512.12±.1836.964±.0299.334±.079
T2M0.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
MDM (ours)0.396±.0040.497±.0219.191±.02210.847±.1091.907±.214
", + "type": "table", + "image_path": "b4c373214b840b1b1680f5c0558314e7810c86b9be268ce24730e3e98f5a8f4c.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 108, + 280, + 501, + 312.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 312.0, + 501, + 344.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 344.0, + 501, + 376.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 203, + 384, + 408, + 396 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 203, + 384, + 408, + 397 + ], + "spans": [ + { + "bbox": [ + 203, + 384, + 408, + 397 + ], + "score": 1.0, + "content": "Table 2: Quantitative results on the KIT test set.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "We evaluate our model using the set of metrics suggested by Guo et al. (2020), namely Frechet In- ´", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "ception Distance (FID), action recognition accuracy, diversity and multimodality. The combination", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 446, + 465, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 465, + 459 + ], + "score": 1.0, + "content": "of these metrics makes a good measure of the realism and diversity of generated motions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Data. HumanAct12 (Guo et al., 2020) offers approximately 1200 motion clips, organized into 12", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "action categories, with 47 to 218 samples per label. UESTC (Ji et al., 2018) consists of 40 action", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "classes, 40 subjects and 25K samples, and is split to train and test. We adhere to the cross-subject", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "testing protocol used by current works, with 225-345 samples per action class. For both datasets we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 507, + 323, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 323, + 519 + ], + "score": 1.0, + "content": "use the sequences provided by Petrovich et al. (2021).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 155, + 558, + 445, + 675 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 155, + 558, + 445, + 675 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 558, + 445, + 675 + ], + "spans": [ + { + "bbox": [ + 155, + 558, + 445, + 675 + ], + "score": 0.961, + "type": "image", + "image_path": "d7a9d635bd4c7d3904d77117eb65b18a0c3237c9cb38308379a700f6d61b9f3c.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 155, + 558, + 445, + 597.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 155, + 597.0, + 445, + 636.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 155, + 636.0, + 445, + 675.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 683, + 506, + 739 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "score": 1.0, + "content": "Figure 4: (a) Text-to-motion user study for the KIT dataset. Each bar represents the preference", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 695, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 505, + 707 + ], + "score": 1.0, + "content": "rate of MDM over the compared model. MDM was preferred over the other models in most of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 705, + 505, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 148, + 718 + ], + "score": 1.0, + "content": "time, and", + "type": "text" + }, + { + "bbox": [ + 148, + 705, + 176, + 716 + ], + "score": 0.88, + "content": "4 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 705, + 463, + 718 + ], + "score": 1.0, + "content": "of the cases even over ground truth samples. The dashed line marks", + "type": "text" + }, + { + "bbox": [ + 464, + 705, + 483, + 716 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 705, + 505, + 718 + ], + "score": 1.0, + "content": ". (b)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 715, + 506, + 729 + ], + "spans": [ + { + "bbox": [ + 105, + 715, + 315, + 729 + ], + "score": 1.0, + "content": "Guidance-scale sweep for HumanML3D dataset.", + "type": "text" + }, + { + "bbox": [ + 315, + 716, + 333, + 727 + ], + "score": 0.32, + "content": "F I D", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 715, + 416, + 729 + ], + "score": 1.0, + "content": "(lower is better) and", + "type": "text" + }, + { + "bbox": [ + 416, + 717, + 424, + 726 + ], + "score": 0.54, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 715, + 506, + 729 + ], + "score": 1.0, + "content": "-precision (higher is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 727, + 495, + 740 + ], + "spans": [ + { + "bbox": [ + 106, + 728, + 268, + 740 + ], + "score": 1.0, + "content": "better) metrics as a function of the scale", + "type": "text" + }, + { + "bbox": [ + 268, + 730, + 274, + 737 + ], + "score": 0.5, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 728, + 459, + 740 + ], + "score": 1.0, + "content": ", draws an accuracy-fidelity sweet spot around", + "type": "text" + }, + { + "bbox": [ + 460, + 727, + 491, + 738 + ], + "score": 0.89, + "content": "s = 2 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 728, + 495, + 740 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + } + ], + "index": 22.0 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 501, + 215 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 70, + 501, + 215 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 70, + 501, + 215 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 501, + 215 + ], + "score": 0.984, + "html": "
MethodRPrecision (top 3)↑FID↓Multimodal Dist↓Diversity→Multimodality↑
Real0.797±.0020.002±.0002.974±.0089.503±.065
JL2P0.486±.00211.02±.0465.296±.0087.676±.0581
Text2Gesture0.345±.0027.664±.0306.030±.0086.409±.0711
T2M0.740±.0031.067±.0023.340±.0089.188±.0022.090±.083
MDM (ours)0.611±.0070.544±.0445.566±.0279.559±.0862.799±.072
+ sent-BERT0.609±.0060.586±.0365.504±.039.666±.0952.707±.188
MDM (decoder)0.608±.0050.767±.0855.507±.0209.176±.0702.927±.125
+ input token0.621±.0050.567±.0515.424±.0229.425±.0602.834±.095
MDM (U-net)0.603±.0061.137±.0085.629±.0328.958±.0982.636±.214
MDM (GRU)0.645±.0054.569±.1505.325±.0267.688±.0821.2646±.024
", + "type": "table", + "image_path": "a8390e456fc75cdd5ea1504b7990139329e00b053fc9fa17c46ccd72412e5cff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 70, + 501, + 118.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 118.33333333333334, + 501, + 166.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 166.66666666666669, + 501, + 215.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 223, + 505, + 268 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "Table 1: Quantitative results on the HumanML3D test set. All methods use the real motion length", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 235, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 201, + 246 + ], + "score": 1.0, + "content": "from the ground truth.", + "type": "text" + }, + { + "bbox": [ + 201, + 235, + 220, + 245 + ], + "score": 0.82, + "content": "\\cdot _ { } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 235, + 505, + 246 + ], + "score": 1.0, + "content": "means results are better if the metric is closer to the real distribution.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 245, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 418, + 258 + ], + "score": 1.0, + "content": "We run all the evaluation 20 times (except MultiModality runs 5 times) and", + "type": "text" + }, + { + "bbox": [ + 418, + 246, + 429, + 256 + ], + "score": 0.67, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 245, + 484, + 258 + ], + "score": 1.0, + "content": "indicates the", + "type": "text" + }, + { + "bbox": [ + 484, + 245, + 504, + 256 + ], + "score": 0.85, + "content": "9 5 \\%", + "type": "inline_equation" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 257, + 293, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 293, + 268 + ], + "score": 1.0, + "content": "confidence interval. Bold indicates best result.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 108, + 280, + 501, + 376 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 280, + 501, + 376 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 280, + 501, + 376 + ], + "spans": [ + { + "bbox": [ + 108, + 280, + 501, + 376 + ], + "score": 0.983, + "html": "
MethodR Precision (top 3)↑FID↓Multimodal Dist↓Diversity→Multimodality↑
Real0.779±.0060.031±.0042.788±.01211.08±.097-
JL2P0.483±.0056.545±.0725.147±.0309.073±.100-
Text2Gesture0.338±.00512.12±.1836.964±.0299.334±.079
T2M0.693±.0072.770±.1093.401±.00810.91±.1191.482±.065
MDM (ours)0.396±.0040.497±.0219.191±.02210.847±.1091.907±.214
", + "type": "table", + "image_path": "b4c373214b840b1b1680f5c0558314e7810c86b9be268ce24730e3e98f5a8f4c.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 108, + 280, + 501, + 312.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 312.0, + 501, + 344.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 344.0, + 501, + 376.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 203, + 384, + 408, + 396 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 203, + 384, + 408, + 397 + ], + "spans": [ + { + "bbox": [ + 203, + 384, + 408, + 397 + ], + "score": 1.0, + "content": "Table 2: Quantitative results on the KIT test set.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "We evaluate our model using the set of metrics suggested by Guo et al. (2020), namely Frechet In- ´", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "ception Distance (FID), action recognition accuracy, diversity and multimodality. The combination", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 446, + 465, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 465, + 459 + ], + "score": 1.0, + "content": "of these metrics makes a good measure of the realism and diversity of generated motions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 424, + 505, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Data. HumanAct12 (Guo et al., 2020) offers approximately 1200 motion clips, organized into 12", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "action categories, with 47 to 218 samples per label. UESTC (Ji et al., 2018) consists of 40 action", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "classes, 40 subjects and 25K samples, and is split to train and test. We adhere to the cross-subject", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "testing protocol used by current works, with 225-345 samples per action class. For both datasets we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 507, + 323, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 323, + 519 + ], + "score": 1.0, + "content": "use the sequences provided by Petrovich et al. (2021).", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 462, + 506, + 519 + ] + }, + { + "type": "image", + "bbox": [ + 155, + 558, + 445, + 675 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 155, + 558, + 445, + 675 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 155, + 558, + 445, + 675 + ], + "spans": [ + { + "bbox": [ + 155, + 558, + 445, + 675 + ], + "score": 0.961, + "type": "image", + "image_path": "d7a9d635bd4c7d3904d77117eb65b18a0c3237c9cb38308379a700f6d61b9f3c.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 155, + 558, + 445, + 597.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 155, + 597.0, + 445, + 636.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 155, + 636.0, + 445, + 675.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 683, + 506, + 739 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 505, + 695 + ], + "score": 1.0, + "content": "Figure 4: (a) Text-to-motion user study for the KIT dataset. Each bar represents the preference", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 695, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 505, + 707 + ], + "score": 1.0, + "content": "rate of MDM over the compared model. MDM was preferred over the other models in most of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 705, + 505, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 148, + 718 + ], + "score": 1.0, + "content": "time, and", + "type": "text" + }, + { + "bbox": [ + 148, + 705, + 176, + 716 + ], + "score": 0.88, + "content": "4 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 705, + 463, + 718 + ], + "score": 1.0, + "content": "of the cases even over ground truth samples. The dashed line marks", + "type": "text" + }, + { + "bbox": [ + 464, + 705, + 483, + 716 + ], + "score": 0.88, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 705, + 505, + 718 + ], + "score": 1.0, + "content": ". (b)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 715, + 506, + 729 + ], + "spans": [ + { + "bbox": [ + 105, + 715, + 315, + 729 + ], + "score": 1.0, + "content": "Guidance-scale sweep for HumanML3D dataset.", + "type": "text" + }, + { + "bbox": [ + 315, + 716, + 333, + 727 + ], + "score": 0.32, + "content": "F I D", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 715, + 416, + 729 + ], + "score": 1.0, + "content": "(lower is better) and", + "type": "text" + }, + { + "bbox": [ + 416, + 717, + 424, + 726 + ], + "score": 0.54, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 715, + 506, + 729 + ], + "score": 1.0, + "content": "-precision (higher is", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 727, + 495, + 740 + ], + "spans": [ + { + "bbox": [ + 106, + 728, + 268, + 740 + ], + "score": 1.0, + "content": "better) metrics as a function of the scale", + "type": "text" + }, + { + "bbox": [ + 268, + 730, + 274, + 737 + ], + "score": 0.5, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 728, + 459, + 740 + ], + "score": 1.0, + "content": ", draws an accuracy-fidelity sweet spot around", + "type": "text" + }, + { + "bbox": [ + 460, + 727, + 491, + 738 + ], + "score": 0.89, + "content": "s = 2 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 728, + 495, + 740 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + } + ], + "index": 22.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 80, + 495, + 208 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 118, + 80, + 495, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 80, + 495, + 208 + ], + "spans": [ + { + "bbox": [ + 118, + 80, + 495, + 208 + ], + "score": 0.982, + "html": "
MethodFID↓Accuracy↑Diversity→Multimodality→
Real (INR)0.020±.0100.997±.0016.850±.0502.450±.040
Real (ours)0.050±.0000.990±.0006.880±.0202.590±.010
Action2Motion (2020)0.338±.0150.917±.0036.879±.0662.511±.023
ACTOR (2021)0.120±.0000.955±.0086.840±.0302.530±.020
INR (2022)0.088±.0040.973±.0016.881±.0482.569±.040
MDM (ours)0.100±.0000.990±.0006.860±.0502.520±.010
w/o foot contact0.080±.0000.990±.0006.810±.0102.580±.010
", + "type": "table", + "image_path": "71e7b9aeb113ee6e6b17585a4c1630b39f58d6ffe56a9d3d7cc375c2ed3f08c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 80, + 495, + 122.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 122.66666666666666, + 495, + 165.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 165.33333333333331, + 495, + 207.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 216, + 505, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "Table 3: Evaluation of action-to-motion on the HumanAct12 dataset. Our model leads the board", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 228, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 239 + ], + "score": 1.0, + "content": "in three out of four metrics. Ground-truth evaluation results are slightly different for each of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "works, due to implementation differences, such as python package versions. It is important to as-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "score": 1.0, + "content": "sess the diversity and multimodality of each model using its own ground-truth results, as they are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "measured by their distance from GT. We show the GT metrics measured by our model and by the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "leading compared work, INR (Cervantes et al., 2022). Bold indicates best result, underline indicates", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 282, + 466, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 157, + 294 + ], + "score": 1.0, + "content": "second best,", + "type": "text" + }, + { + "bbox": [ + 158, + 283, + 167, + 293 + ], + "score": 0.73, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 282, + 205, + 294 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 205, + 282, + 226, + 293 + ], + "score": 0.85, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 282, + 307, + 294 + ], + "score": 1.0, + "content": "confidence interval,", + "type": "text" + }, + { + "bbox": [ + 307, + 284, + 320, + 293 + ], + "score": 0.82, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 282, + 466, + 294 + ], + "score": 1.0, + "content": "indicates that closer to real is better.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 108, + 306, + 505, + 398 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 306, + 505, + 398 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 306, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 108, + 306, + 505, + 398 + ], + "score": 0.976, + "html": "
MethodFIDtrain FIDtestAccuracy↑Diversity→Multimodality→
Real2.92±.262.79±.290.988±.00133.34±.32014.16±.06
ACTOR (2021)20.49±2.3123.43±2.200.911±.00331.96±.3314.52±.09
INR (2022) (best variation)9.55±.0615.00±.090.941±.00131.59±.1914.68±.07
MDM (ours)9.98±1.3312.81±1.460.950±.00033.02±.2814.26±.12
w/o foot contact9.69±.8113.08±2.320.960±.00033.10±.2914.06±.05
", + "type": "table", + "image_path": "ec3d68c861da7d767ff2deaf102626966cf953b43a64342352012a725813007f.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 306, + 505, + 336.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 336.6666666666667, + 505, + 367.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 367.33333333333337, + 505, + 398.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 406, + 505, + 439 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "Table 4: Evaluation of action-to-motion on the UESTC dataset. The performance improvement", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "with our model shows a clear gap from state-of-the-art. Bold indicates best result, underline indi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 428, + 488, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 180, + 440 + ], + "score": 1.0, + "content": "cates second best,", + "type": "text" + }, + { + "bbox": [ + 180, + 429, + 190, + 439 + ], + "score": 0.76, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 429, + 228, + 440 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 228, + 428, + 248, + 439 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 429, + 329, + 440 + ], + "score": 1.0, + "content": "confidence interval,", + "type": "text" + }, + { + "bbox": [ + 329, + 429, + 342, + 438 + ], + "score": 0.83, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 429, + 488, + 440 + ], + "score": 1.0, + "content": "indicates that closer to real is better.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "Implementation. The implementation presented in Figure 2 holds for all the variations of our work.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "In the case of action-to-motion, the only change would be the substitution of the text embedding by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 485, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 496 + ], + "score": 1.0, + "content": "an action embedding. Since action is represented by a scalar, its embedding is fairly simple; each", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 496, + 478, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 478, + 507 + ], + "score": 1.0, + "content": "input action class scalar is converted into a learned embedding of the transformer dimension.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "The experiments have been run with batch size 64, a latent dimension of 512, and an encoder-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 461, + 534 + ], + "score": 1.0, + "content": "transformer architecture. Training on HumanAct12 and UESTC has been carried out for", + "type": "text" + }, + { + "bbox": [ + 461, + 523, + 487, + 534 + ], + "score": 0.85, + "content": "7 5 0 K", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 123, + 544 + ], + "score": 0.66, + "content": "2 M", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "steps respectively. In our tables we display the evaluation of the checkpoint that minimizes the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 545, + 156, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 156, + 557 + ], + "score": 1.0, + "content": "FID metric.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 561, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Quantitative evaluation. Tables 3 and 4 reflect MDM’s performance on the HumanAct12 and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "UESTC datasets respectively. We conduct 20 evaluations, with 1000 samples in each, and report", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 584, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 185, + 595 + ], + "score": 1.0, + "content": "their average and a", + "type": "text" + }, + { + "bbox": [ + 185, + 584, + 205, + 594 + ], + "score": 0.88, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 584, + 505, + 595 + ], + "score": 1.0, + "content": "confidence interval. We test two variations, with and without foot contact", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 595, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 607 + ], + "score": 1.0, + "content": "loss. Full ablation study for geometric losses is presented in Appendix A.2. Our model leads the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "board for both datasets. The variation with no foot contact loss attains slightly better results; never-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 617, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 629 + ], + "score": 1.0, + "content": "theless, as shown in our supplementary video, the contribution of foot contact loss to the quality of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 628, + 493, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 493, + 639 + ], + "score": 1.0, + "content": "results is important, and without it we witness artifacts such as shakiness and unnatural gestures.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 651, + 271, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 273, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 273, + 666 + ], + "score": 1.0, + "content": "5 ADDITIONAL APPLICATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 208, + 683 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 209, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 209, + 684 + ], + "score": 1.0, + "content": "5.1 MOTION EDITING", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "In this section we implement two motion editing applications - in-betweening and body part edit-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "ing, both using the same approach in the temporal and spatial domains correspondingly. For in-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 250, + 723 + ], + "score": 1.0, + "content": "betweening, we fix the first and last", + "type": "text" + }, + { + "bbox": [ + 251, + 709, + 270, + 720 + ], + "score": 0.88, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "of the motion, leaving the model to generate the remaining", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 126, + 731 + ], + "score": 0.87, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "in the middle. For body part editing, we fix the joints we don’t want to edit and leave the", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 80, + 495, + 208 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 118, + 80, + 495, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 80, + 495, + 208 + ], + "spans": [ + { + "bbox": [ + 118, + 80, + 495, + 208 + ], + "score": 0.982, + "html": "
MethodFID↓Accuracy↑Diversity→Multimodality→
Real (INR)0.020±.0100.997±.0016.850±.0502.450±.040
Real (ours)0.050±.0000.990±.0006.880±.0202.590±.010
Action2Motion (2020)0.338±.0150.917±.0036.879±.0662.511±.023
ACTOR (2021)0.120±.0000.955±.0086.840±.0302.530±.020
INR (2022)0.088±.0040.973±.0016.881±.0482.569±.040
MDM (ours)0.100±.0000.990±.0006.860±.0502.520±.010
w/o foot contact0.080±.0000.990±.0006.810±.0102.580±.010
", + "type": "table", + "image_path": "71e7b9aeb113ee6e6b17585a4c1630b39f58d6ffe56a9d3d7cc375c2ed3f08c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 80, + 495, + 122.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 122.66666666666666, + 495, + 165.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 165.33333333333331, + 495, + 207.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 216, + 505, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "Table 3: Evaluation of action-to-motion on the HumanAct12 dataset. Our model leads the board", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 228, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 505, + 239 + ], + "score": 1.0, + "content": "in three out of four metrics. Ground-truth evaluation results are slightly different for each of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 252 + ], + "score": 1.0, + "content": "works, due to implementation differences, such as python package versions. It is important to as-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "score": 1.0, + "content": "sess the diversity and multimodality of each model using its own ground-truth results, as they are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "measured by their distance from GT. We show the GT metrics measured by our model and by the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "leading compared work, INR (Cervantes et al., 2022). Bold indicates best result, underline indicates", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 282, + 466, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 157, + 294 + ], + "score": 1.0, + "content": "second best,", + "type": "text" + }, + { + "bbox": [ + 158, + 283, + 167, + 293 + ], + "score": 0.73, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 282, + 205, + 294 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 205, + 282, + 226, + 293 + ], + "score": 0.85, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 282, + 307, + 294 + ], + "score": 1.0, + "content": "confidence interval,", + "type": "text" + }, + { + "bbox": [ + 307, + 284, + 320, + 293 + ], + "score": 0.82, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 282, + 466, + 294 + ], + "score": 1.0, + "content": "indicates that closer to real is better.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 108, + 306, + 505, + 398 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 306, + 505, + 398 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 306, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 108, + 306, + 505, + 398 + ], + "score": 0.976, + "html": "
MethodFIDtrain FIDtestAccuracy↑Diversity→Multimodality→
Real2.92±.262.79±.290.988±.00133.34±.32014.16±.06
ACTOR (2021)20.49±2.3123.43±2.200.911±.00331.96±.3314.52±.09
INR (2022) (best variation)9.55±.0615.00±.090.941±.00131.59±.1914.68±.07
MDM (ours)9.98±1.3312.81±1.460.950±.00033.02±.2814.26±.12
w/o foot contact9.69±.8113.08±2.320.960±.00033.10±.2914.06±.05
", + "type": "table", + "image_path": "ec3d68c861da7d767ff2deaf102626966cf953b43a64342352012a725813007f.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 306, + 505, + 336.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 336.6666666666667, + 505, + 367.33333333333337 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 367.33333333333337, + 505, + 398.00000000000006 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 406, + 505, + 439 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "Table 4: Evaluation of action-to-motion on the UESTC dataset. The performance improvement", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 430 + ], + "score": 1.0, + "content": "with our model shows a clear gap from state-of-the-art. Bold indicates best result, underline indi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 428, + 488, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 180, + 440 + ], + "score": 1.0, + "content": "cates second best,", + "type": "text" + }, + { + "bbox": [ + 180, + 429, + 190, + 439 + ], + "score": 0.76, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 429, + 228, + 440 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 228, + 428, + 248, + 439 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 429, + 329, + 440 + ], + "score": 1.0, + "content": "confidence interval,", + "type": "text" + }, + { + "bbox": [ + 329, + 429, + 342, + 438 + ], + "score": 0.83, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 429, + 488, + 440 + ], + "score": 1.0, + "content": "indicates that closer to real is better.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "Implementation. The implementation presented in Figure 2 holds for all the variations of our work.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 505, + 486 + ], + "score": 1.0, + "content": "In the case of action-to-motion, the only change would be the substitution of the text embedding by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 485, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 496 + ], + "score": 1.0, + "content": "an action embedding. Since action is represented by a scalar, its embedding is fairly simple; each", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 496, + 478, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 478, + 507 + ], + "score": 1.0, + "content": "input action class scalar is converted into a learned embedding of the transformer dimension.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 462, + 506, + 507 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "The experiments have been run with batch size 64, a latent dimension of 512, and an encoder-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 461, + 534 + ], + "score": 1.0, + "content": "transformer architecture. Training on HumanAct12 and UESTC has been carried out for", + "type": "text" + }, + { + "bbox": [ + 461, + 523, + 487, + 534 + ], + "score": 0.85, + "content": "7 5 0 K", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 123, + 544 + ], + "score": 0.66, + "content": "2 M", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "steps respectively. In our tables we display the evaluation of the checkpoint that minimizes the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 545, + 156, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 156, + 557 + ], + "score": 1.0, + "content": "FID metric.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 512, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 561, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "Quantitative evaluation. Tables 3 and 4 reflect MDM’s performance on the HumanAct12 and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "UESTC datasets respectively. We conduct 20 evaluations, with 1000 samples in each, and report", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 584, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 185, + 595 + ], + "score": 1.0, + "content": "their average and a", + "type": "text" + }, + { + "bbox": [ + 185, + 584, + 205, + 594 + ], + "score": 0.88, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 584, + 505, + 595 + ], + "score": 1.0, + "content": "confidence interval. We test two variations, with and without foot contact", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 595, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 607 + ], + "score": 1.0, + "content": "loss. Full ablation study for geometric losses is presented in Appendix A.2. Our model leads the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 506, + 618 + ], + "score": 1.0, + "content": "board for both datasets. The variation with no foot contact loss attains slightly better results; never-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 617, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 506, + 629 + ], + "score": 1.0, + "content": "theless, as shown in our supplementary video, the contribution of foot contact loss to the quality of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 628, + 493, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 493, + 639 + ], + "score": 1.0, + "content": "results is important, and without it we witness artifacts such as shakiness and unnatural gestures.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 561, + 506, + 639 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 651, + 271, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 273, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 273, + 666 + ], + "score": 1.0, + "content": "5 ADDITIONAL APPLICATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 208, + 683 + ], + "lines": [ + { + "bbox": [ + 106, + 671, + 209, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 209, + 684 + ], + "score": 1.0, + "content": "5.1 MOTION EDITING", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "In this section we implement two motion editing applications - in-betweening and body part edit-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "ing, both using the same approach in the temporal and spatial domains correspondingly. For in-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 250, + 723 + ], + "score": 1.0, + "content": "betweening, we fix the first and last", + "type": "text" + }, + { + "bbox": [ + 251, + 709, + 270, + 720 + ], + "score": 0.88, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "of the motion, leaving the model to generate the remaining", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 126, + 731 + ], + "score": 0.87, + "content": "5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "in the middle. For body part editing, we fix the joints we don’t want to edit and leave the", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 126, + 472, + 194 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 82, + 505, + 117 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "model to generate the rest. In particular, we experiment with editing the upper body joints only.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "In figure 3 we show that in both cases, using the method described in Section 3 generates smooth", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 477, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 477, + 117 + ], + "score": 1.0, + "content": "motions that adhere both to the fixed part of the motion and the condition (if one was given).", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 139, + 126, + 472, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 139, + 126, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 139, + 126, + 472, + 194 + ], + "score": 0.98, + "html": "
MethodFID↓KID↓Precision↑ Recall↑Diversity↑
ACTOR (2021)48.800.530.72, 0.7414.10
MoDi (2022)13.030.120.71, 0.8117.57
MDM (ours)31.920.360.66,0.6217.00
", + "type": "table", + "image_path": "089bb7a3f1346cccaada3c6555301bd17b789d1dbb439201a15aa8001fc8c5cb.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 139, + 126, + 472, + 148.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 139, + 148.66666666666666, + 472, + 171.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 139, + 171.33333333333331, + 472, + 193.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 202, + 504, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "score": 1.0, + "content": "Table 5: Evaluation of unconstrained synthesis on the HumanAct12 dataset. We test MDM in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "the challenging unconstrained setting, and compare with MoDi (Raab et al., 2022), a work that was", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 224, + 504, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 504, + 237 + ], + "score": 1.0, + "content": "specially designed for such setting. We demonstrate that in addition to being able to support any", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 235, + 500, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 500, + 247 + ], + "score": 1.0, + "content": "condition, we can achieve plausible results in the unconstrained setting. Bold indicates best result.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 252, + 258, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 259, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 259, + 264 + ], + "score": 1.0, + "content": "5.2 UNCONSTRAINED SYNTHESIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "The challenging task of unconstrained synthesis has been studied by only a few (Holden et al., 2016;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Raab et al., 2022). In the presence of data labeling, e.g., action classes or text description, the la-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 290, + 504, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 504, + 302 + ], + "score": 1.0, + "content": "bels work as a supervising factor, and facilitate a structured latent space for the training network.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "The lack of labeling make training more difficult. The human motion field possesses rich unlabeled", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "datasets (Adobe Systems Inc., 2021), and the ability to train on top of them is an advantage. Daring", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 323, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 334 + ], + "score": 1.0, + "content": "to test MDM in the challenging unconstrained setting, we follow MoDi(Raab et al., 2022) for eval-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "uation. We use the metrics they suggest (FID, KID, precision/recall and multimodality), and run on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 345, + 396, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 396, + 357 + ], + "score": 1.0, + "content": "an unconstrained version of the HumanAct12 (Guo et al., 2020) dataset.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Data. Although annotated, we use HumanAct12 (see Section 4.2) in an unconstrained fashion,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "ignoring its labels. The choice of HumanAct12 rather than a dataset with no labels (e.g., Mix-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 436, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 436, + 396 + ], + "score": 1.0, + "content": "amo (Adobe Systems Inc., 2021)), is for compatibility with previous publications.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Implementation. Our model uses the same architecture for all forms of conditioning, as well as for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "the unconstrained setting. The only change to the structure shown in Figure 2, is the removal of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 215, + 435 + ], + "score": 1.0, + "content": "conditional input, such that", + "type": "text" + }, + { + "bbox": [ + 216, + 423, + 230, + 433 + ], + "score": 0.88, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 421, + 359, + 435 + ], + "score": 1.0, + "content": "is composed of the projection of", + "type": "text" + }, + { + "bbox": [ + 359, + 423, + 364, + 432 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "only. To simulate an unconstrained", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "behavior, ACTOR Petrovich et al. (2021) has been trained by (Raab et al., 2022) with a labeling of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 205, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 205, + 456 + ], + "score": 1.0, + "content": "one class to all motions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "Quantitative evaluation. The results of our evaluation are shown in table 5. We demonstrate supe-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "score": 1.0, + "content": "riority over works that were not designed for an unconstrained setting, and get closer to MoDi (Raab", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "et al., 2022). MoDi is carefully molded for unconstrained settings, while our work can be applied to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 495, + 347, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 347, + 506 + ], + "score": 1.0, + "content": "any (or no) constrain, and also provides editing capabilities.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "title", + "bbox": [ + 107, + 516, + 190, + 529 + ], + "lines": [ + { + "bbox": [ + 104, + 514, + 192, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 192, + 532 + ], + "score": 1.0, + "content": "6 DISCUSSION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "score": 1.0, + "content": "We have presented MDM, a method that lends itself to various human motion generation tasks.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "score": 1.0, + "content": "MDM is an untypical classifier-free diffusion model, featuring a transformer-encoder backbone, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "predicting the signal, rather than the noise. This yields both a lightweight model, that is unburdening", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "to train, and an accurate one, gaining much from the applicable geometric losses. Our experiments", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "show superiority in conditioned generation, but also that this approach is not very sensitive to the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "choice of architecture. A notable limitation of the diffusion approach is the long inference time,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "score": 1.0, + "content": "requiring about 1000 forward passes for a single result. Since our motion model is small anyway,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 613, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 625 + ], + "score": 1.0, + "content": "using dimensions order of magnitude smaller than images, our inference time shifts from less than", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "score": 1.0, + "content": "a second to only about a minute, which is an acceptable compromise. As diffusion models continue", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "to evolve, besides better compute, in the future we would be interested in seeing how to incorporate", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 646, + 479, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 479, + 659 + ], + "score": 1.0, + "content": "better control into the generation process and widen the options for applications even further.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37 + }, + { + "type": "title", + "bbox": [ + 108, + 675, + 225, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 226, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 226, + 689 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "We thank Rinon Gal for his useful suggestions and references. This research was supported in part", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "by the Israel Science Foundation (grants no. 2492/20 and 3441/21), Len Blavatnik and the Blavatnik", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 436, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 436, + 732 + ], + "score": 1.0, + "content": "family foundation, and The Tel Aviv University Innovation Laboratories (TILabs).", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 139, + 126, + 472, + 194 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 82, + 505, + 117 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "model to generate the rest. In particular, we experiment with editing the upper body joints only.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "In figure 3 we show that in both cases, using the method described in Section 3 generates smooth", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 477, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 477, + 117 + ], + "score": 1.0, + "content": "motions that adhere both to the fixed part of the motion and the condition (if one was given).", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 139, + 126, + 472, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 139, + 126, + 472, + 194 + ], + "spans": [ + { + "bbox": [ + 139, + 126, + 472, + 194 + ], + "score": 0.98, + "html": "
MethodFID↓KID↓Precision↑ Recall↑Diversity↑
ACTOR (2021)48.800.530.72, 0.7414.10
MoDi (2022)13.030.120.71, 0.8117.57
MDM (ours)31.920.360.66,0.6217.00
", + "type": "table", + "image_path": "089bb7a3f1346cccaada3c6555301bd17b789d1dbb439201a15aa8001fc8c5cb.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 139, + 126, + 472, + 148.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 139, + 148.66666666666666, + 472, + 171.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 139, + 171.33333333333331, + 472, + 193.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 202, + 504, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "score": 1.0, + "content": "Table 5: Evaluation of unconstrained synthesis on the HumanAct12 dataset. We test MDM in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "the challenging unconstrained setting, and compare with MoDi (Raab et al., 2022), a work that was", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 224, + 504, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 504, + 237 + ], + "score": 1.0, + "content": "specially designed for such setting. We demonstrate that in addition to being able to support any", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 235, + 500, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 500, + 247 + ], + "score": 1.0, + "content": "condition, we can achieve plausible results in the unconstrained setting. Bold indicates best result.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 252, + 258, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 251, + 259, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 259, + 264 + ], + "score": 1.0, + "content": "5.2 UNCONSTRAINED SYNTHESIS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "The challenging task of unconstrained synthesis has been studied by only a few (Holden et al., 2016;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "Raab et al., 2022). In the presence of data labeling, e.g., action classes or text description, the la-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 290, + 504, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 504, + 302 + ], + "score": 1.0, + "content": "bels work as a supervising factor, and facilitate a structured latent space for the training network.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 505, + 313 + ], + "score": 1.0, + "content": "The lack of labeling make training more difficult. The human motion field possesses rich unlabeled", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "datasets (Adobe Systems Inc., 2021), and the ability to train on top of them is an advantage. Daring", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 323, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 334 + ], + "score": 1.0, + "content": "to test MDM in the challenging unconstrained setting, we follow MoDi(Raab et al., 2022) for eval-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "uation. We use the metrics they suggest (FID, KID, precision/recall and multimodality), and run on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 345, + 396, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 396, + 357 + ], + "score": 1.0, + "content": "an unconstrained version of the HumanAct12 (Guo et al., 2020) dataset.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 268, + 505, + 357 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Data. Although annotated, we use HumanAct12 (see Section 4.2) in an unconstrained fashion,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "ignoring its labels. The choice of HumanAct12 rather than a dataset with no labels (e.g., Mix-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 383, + 436, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 436, + 396 + ], + "score": 1.0, + "content": "amo (Adobe Systems Inc., 2021)), is for compatibility with previous publications.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 361, + 505, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Implementation. Our model uses the same architecture for all forms of conditioning, as well as for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "the unconstrained setting. The only change to the structure shown in Figure 2, is the removal of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 215, + 435 + ], + "score": 1.0, + "content": "conditional input, such that", + "type": "text" + }, + { + "bbox": [ + 216, + 423, + 230, + 433 + ], + "score": 0.88, + "content": "z _ { t k }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 421, + 359, + 435 + ], + "score": 1.0, + "content": "is composed of the projection of", + "type": "text" + }, + { + "bbox": [ + 359, + 423, + 364, + 432 + ], + "score": 0.69, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "only. To simulate an unconstrained", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 506, + 446 + ], + "score": 1.0, + "content": "behavior, ACTOR Petrovich et al. (2021) has been trained by (Raab et al., 2022) with a labeling of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 444, + 205, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 205, + 456 + ], + "score": 1.0, + "content": "one class to all motions.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 400, + 506, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 474 + ], + "score": 1.0, + "content": "Quantitative evaluation. The results of our evaluation are shown in table 5. We demonstrate supe-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 506, + 484 + ], + "score": 1.0, + "content": "riority over works that were not designed for an unconstrained setting, and get closer to MoDi (Raab", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 496 + ], + "score": 1.0, + "content": "et al., 2022). MoDi is carefully molded for unconstrained settings, while our work can be applied to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 495, + 347, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 347, + 506 + ], + "score": 1.0, + "content": "any (or no) constrain, and also provides editing capabilities.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 460, + 506, + 506 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 516, + 190, + 529 + ], + "lines": [ + { + "bbox": [ + 104, + 514, + 192, + 532 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 192, + 532 + ], + "score": 1.0, + "content": "6 DISCUSSION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 504, + 549 + ], + "score": 1.0, + "content": "We have presented MDM, a method that lends itself to various human motion generation tasks.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 505, + 559 + ], + "score": 1.0, + "content": "MDM is an untypical classifier-free diffusion model, featuring a transformer-encoder backbone, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "predicting the signal, rather than the noise. This yields both a lightweight model, that is unburdening", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "to train, and an accurate one, gaining much from the applicable geometric losses. Our experiments", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "show superiority in conditioned generation, but also that this approach is not very sensitive to the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "choice of architecture. A notable limitation of the diffusion approach is the long inference time,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "score": 1.0, + "content": "requiring about 1000 forward passes for a single result. Since our motion model is small anyway,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 613, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 625 + ], + "score": 1.0, + "content": "using dimensions order of magnitude smaller than images, our inference time shifts from less than", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 505, + 636 + ], + "score": 1.0, + "content": "a second to only about a minute, which is an acceptable compromise. As diffusion models continue", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 648 + ], + "score": 1.0, + "content": "to evolve, besides better compute, in the future we would be interested in seeing how to incorporate", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 646, + 479, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 479, + 659 + ], + "score": 1.0, + "content": "better control into the generation process and widen the options for applications even further.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 536, + 506, + 659 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 675, + 225, + 686 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 226, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 226, + 689 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "We thank Rinon Gal for his useful suggestions and references. This research was supported in part", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "by the Israel Science Foundation (grants no. 2492/20 and 3441/21), Len Blavatnik and the Blavatnik", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 436, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 436, + 732 + ], + "score": 1.0, + "content": "family foundation, and The Tel Aviv University Innovation Laboratories (TILabs).", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 504, + 134 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 222, + 134 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 222, + 134 + ], + "score": 1.0, + "content": "(TOG), 39(4):62–1, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 105, + 141, + 504, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 154 + ], + "score": 1.0, + "content": "Adobe Systems Inc. Mixamo, 2021. URL https://www.mixamo.com. Accessed: 2021-12-25.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 160, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 106, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 171, + 494, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 494, + 185 + ], + "score": 1.0, + "content": "forecasting. In 2019 International Conference on 3D Vision (3DV), pp. 719–728. IEEE, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 504, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pp. 565–574.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 212, + 169, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 212, + 169, + 226 + ], + "score": 1.0, + "content": "IEEE, 2021.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 232, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "A Aristidou, A Yiannakidis, K Aberman, D Cohen-Or, A Shamir, and Y Chrysanthou. Rhythm is a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 244, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 117, + 244, + 505, + 255 + ], + "score": 1.0, + "content": "dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 255, + 243, + 267 + ], + "spans": [ + { + "bbox": [ + 116, + 255, + 243, + 267 + ], + "score": 1.0, + "content": "and Computer Graphics, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body ges-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "tures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pp. 1–10.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 307, + 169, + 319 + ], + "spans": [ + { + "bbox": [ + 115, + 307, + 169, + 319 + ], + "score": 1.0, + "content": "IEEE, 2021.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 504, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 337, + 463, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 337, + 463, + 350 + ], + "score": 1.0, + "content": "for variable length human motion generation. arXiv preprint arXiv:2203.13694, 2022.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 506, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 368, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 506, + 380 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 378, + 412, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 412, + 391 + ], + "score": 1.0, + "content": "for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 409, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 116, + 409, + 506, + 422 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 419, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 506, + 434 + ], + "score": 1.0, + "content": "the North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, Minneapolis, Minnesota, June", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 440, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 115, + 440, + 505, + 456 + ], + "score": 1.0, + "content": "2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URL https:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 117, + 453, + 284, + 465 + ], + "spans": [ + { + "bbox": [ + 117, + 453, + 284, + 465 + ], + "score": 1.0, + "content": "//aclanthology.org/N19-1423.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 504, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "score": 1.0, + "content": "Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 483, + 378, + 496 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 378, + 496 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, 34:8780–8794, 2021.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 513, + 477, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 513, + 477, + 526 + ], + "score": 1.0, + "content": "Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "for human dynamics. In Proceedings of the IEEE international conference on computer vision,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 556, + 206, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 556, + 206, + 567 + ], + "score": 1.0, + "content": "pp. 4346–4354, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 506, + 588 + ], + "score": 1.0, + "content": "Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 585, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 505, + 599 + ], + "score": 1.0, + "content": "Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 596, + 493, + 610 + ], + "spans": [ + { + "bbox": [ + 115, + 596, + 493, + 610 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17113–17122, 2022.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 615, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 637, + 415, + 651 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 415, + 651 + ], + "score": 1.0, + "content": "28th ACM International Conference on Multimedia, pp. 2021–2029, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 672 + ], + "score": 1.0, + "content": "Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 678, + 383, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 383, + 693 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition, pp. 5152–5161, 2022a.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Noguer. Back to mlp: A simple baseline for human motion prediction. arXiv preprint", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 721, + 224, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 224, + 731 + ], + "score": 1.0, + "content": "arXiv:2207.01567, 2022b.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 504, + 134 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 222, + 134 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 222, + 134 + ], + "score": 1.0, + "content": "(TOG), 39(4):62–1, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 99, + 505, + 134 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 141, + 504, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 140, + 505, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 140, + 505, + 154 + ], + "score": 1.0, + "content": "Adobe Systems Inc. Mixamo, 2021. URL https://www.mixamo.com. Accessed: 2021-12-25.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 140, + 505, + 154 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 160, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 106, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 171, + 494, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 494, + 185 + ], + "score": 1.0, + "content": "forecasting. In 2019 International Conference on 3D Vision (3DV), pp. 719–728. IEEE, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 160, + 505, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 504, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 202, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 506, + 216 + ], + "score": 1.0, + "content": "3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pp. 565–574.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 212, + 169, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 212, + 169, + 226 + ], + "score": 1.0, + "content": "IEEE, 2021.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 191, + 506, + 226 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 232, + 505, + 266 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "A Aristidou, A Yiannakidis, K Aberman, D Cohen-Or, A Shamir, and Y Chrysanthou. Rhythm is a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 244, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 117, + 244, + 505, + 255 + ], + "score": 1.0, + "content": "dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 255, + 243, + 267 + ], + "spans": [ + { + "bbox": [ + 116, + 255, + 243, + 267 + ], + "score": 1.0, + "content": "and Computer Graphics, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 232, + 506, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body ges-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 115, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "tures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pp. 1–10.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 307, + 169, + 319 + ], + "spans": [ + { + "bbox": [ + 115, + 307, + 169, + 319 + ], + "score": 1.0, + "content": "IEEE, 2021.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 275, + 505, + 319 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 504, + 349 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 337, + 463, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 337, + 463, + 350 + ], + "score": 1.0, + "content": "for variable length human motion generation. arXiv preprint arXiv:2203.13694, 2022.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 326, + 505, + 350 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 506, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 368, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 506, + 380 + ], + "score": 1.0, + "content": "ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 378, + 412, + 391 + ], + "spans": [ + { + "bbox": [ + 116, + 378, + 412, + 391 + ], + "score": 1.0, + "content": "for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 356, + 506, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 465 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 506, + 411 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 409, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 116, + 409, + 506, + 422 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 419, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 506, + 434 + ], + "score": 1.0, + "content": "the North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 115, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, Minneapolis, Minnesota, June", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 440, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 115, + 440, + 505, + 456 + ], + "score": 1.0, + "content": "2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URL https:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 117, + 453, + 284, + 465 + ], + "spans": [ + { + "bbox": [ + 117, + 453, + 284, + 465 + ], + "score": 1.0, + "content": "//aclanthology.org/N19-1423.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 397, + 506, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 472, + 504, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 505, + 486 + ], + "score": 1.0, + "content": "Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 483, + 378, + 496 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 378, + 496 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, 34:8780–8794, 2021.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 470, + 505, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 503, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 513, + 477, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 513, + 477, + 526 + ], + "score": 1.0, + "content": "Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 502, + 505, + 526 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 532, + 505, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "for human dynamics. In Proceedings of the IEEE international conference on computer vision,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 556, + 206, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 556, + 206, + 567 + ], + "score": 1.0, + "content": "pp. 4346–4354, 2015.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 533, + 505, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 106, + 574, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 506, + 588 + ], + "score": 1.0, + "content": "Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 585, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 505, + 599 + ], + "score": 1.0, + "content": "Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 596, + 493, + 610 + ], + "spans": [ + { + "bbox": [ + 115, + 596, + 493, + 610 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17113–17122, 2022.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 574, + 506, + 610 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 615, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 637, + 415, + 651 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 415, + 651 + ], + "score": 1.0, + "content": "28th ACM International Conference on Multimedia, pp. 2021–2029, 2020.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 615, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 672 + ], + "score": 1.0, + "content": "Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 678, + 383, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 383, + 693 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition, pp. 5152–5161, 2022a.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 654, + 506, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Noguer. Back to mlp: A simple baseline for human motion prediction. arXiv preprint", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 721, + 224, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 224, + 731 + ], + "score": 1.0, + "content": "arXiv:2207.01567, 2022b.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 698, + 505, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Felix G Harvey and Christopher Pal. Recurrent transition networks for character locomotion. In ´", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 339, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 339, + 105 + ], + "score": 1.0, + "content": "SIGGRAPH Asia 2018 Technical Briefs, pp. 1–4. 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 105, + 111, + 503, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Felix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in- ´", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 403, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 403, + 135 + ], + "score": 1.0, + "content": "betweening. ACM Transactions on Graphics (TOG), 39(4):60–1, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 141, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Com-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 259, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 259, + 176 + ], + "score": 1.0, + "content": "puter Vision, pp. 7134–7143, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 104, + 182, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 258, + 195 + ], + "score": 1.0, + "content": "Jonathan Ho and Tim Salimans.", + "type": "text" + }, + { + "bbox": [ + 273, + 180, + 430, + 196 + ], + "score": 1.0, + "content": "Classifier-free diffusion guidance.", + "type": "text" + }, + { + "bbox": [ + 438, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "arXiv preprint", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 219, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 219, + 205 + ], + "score": 1.0, + "content": "arXiv:2207.12598, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 367, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 367, + 236 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 33:6840–6851, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 241, + 504, + 265 + ], + "lines": [ + { + "bbox": [ + 105, + 242, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 255 + ], + "score": 1.0, + "content": "Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 253, + 404, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 253, + 404, + 266 + ], + "score": 1.0, + "content": "Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 504, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 284, + 440, + 295 + ], + "spans": [ + { + "bbox": [ + 116, + 284, + 440, + 295 + ], + "score": 1.0, + "content": "synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1–11, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 108, + 301, + 503, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "score": 1.0, + "content": "Yanli Ji, Feixiang Xu, Yang Yang, Fumin Shen, Heng Tao Shen, and Wei-Shi Zheng. A large-scale", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "rgb-d database for arbitrary-view human action recognition. In Proceedings of the 26th ACM", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 324, + 372, + 336 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 372, + 336 + ], + "score": 1.0, + "content": "international Conference on Multimedia, pp. 1510–1518, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 108, + 342, + 504, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Con-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "volutional autoencoders for human motion infilling. In 2020 International Conference on 3D", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 117, + 365, + 279, + 377 + ], + "spans": [ + { + "bbox": [ + 117, + 365, + 279, + 377 + ], + "score": 1.0, + "content": "Vision (3DV), pp. 918–927. IEEE, 2020.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 503, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 394, + 322, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 322, + 407 + ], + "score": 1.0, + "content": "& editing. arXiv preprint arXiv:2209.00349, 2022.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 412, + 503, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 146, + 426 + ], + "score": 1.0, + "content": "Diederik", + "type": "text" + }, + { + "bbox": [ + 146, + 414, + 155, + 424 + ], + "score": 0.29, + "content": "\\mathrm { \\bf P }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 424, + 214, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 214, + 435 + ], + "score": 1.0, + "content": "arXiv:1312.6114, 2013.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 504, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 116, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 466, + 305, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 305, + 477 + ], + "score": 1.0, + "content": "and pattern recognition, pp. 5253–5263, 2020.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 504, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 504, + 497 + ], + "score": 1.0, + "content": "Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music con-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 116, + 495, + 272, + 507 + ], + "score": 1.0, + "content": "ditioned 3d dance generation with aist", + "type": "text" + }, + { + "bbox": [ + 273, + 496, + 284, + 505 + ], + "score": 0.38, + "content": "^ { + + }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 495, + 505, + 507 + ], + "score": 1.0, + "content": ". In The IEEE International Conference on Computer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 506, + 203, + 518 + ], + "spans": [ + { + "bbox": [ + 117, + 506, + 203, + 518 + ], + "score": 1.0, + "content": "Vision (ICCV), 2021.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceed-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 115, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2837–2845,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 545, + 143, + 559 + ], + "spans": [ + { + "bbox": [ + 114, + 545, + 143, + 559 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 108, + 564, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 116, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "AMASS: Archive of motion capture as surface shapes. In International Conference on Computer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 117, + 587, + 271, + 599 + ], + "spans": [ + { + "bbox": [ + 117, + 587, + 271, + 599 + ], + "score": 1.0, + "content": "Vision, pp. 5442–5451, October 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recur-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "rent neural networks. In Proceedings of the IEEE conference on computer vision and pattern", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 628, + 257, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 257, + 640 + ], + "score": 1.0, + "content": "recognition, pp. 2891–2900, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 658, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 116, + 658, + 505, + 670 + ], + "score": 1.0, + "content": "Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 669, + 399, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 399, + 681 + ], + "score": 1.0, + "content": "text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "score": 1.0, + "content": "Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dim-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "itrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR),", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 216, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 216, + 732 + ], + "score": 1.0, + "content": "pp. 10975–10985, 2019.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Felix G Harvey and Christopher Pal. Recurrent transition networks for character locomotion. In ´", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 339, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 339, + 105 + ], + "score": 1.0, + "content": "SIGGRAPH Asia 2018 Technical Briefs, pp. 1–4. 2018.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 83, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 111, + 503, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Felix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in- ´", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 123, + 403, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 123, + 403, + 135 + ], + "score": 1.0, + "content": "betweening. ACM Transactions on Graphics (TOG), 39(4):60–1, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 112, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 141, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 155 + ], + "score": 1.0, + "content": "Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Com-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 259, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 259, + 176 + ], + "score": 1.0, + "content": "puter Vision, pp. 7134–7143, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 106, + 141, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 182, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 258, + 195 + ], + "score": 1.0, + "content": "Jonathan Ho and Tim Salimans.", + "type": "text" + }, + { + "bbox": [ + 273, + 180, + 430, + 196 + ], + "score": 1.0, + "content": "Classifier-free diffusion guidance.", + "type": "text" + }, + { + "bbox": [ + 438, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "arXiv preprint", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 219, + 205 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 219, + 205 + ], + "score": 1.0, + "content": "arXiv:2207.12598, 2022.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 180, + 506, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 212, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 223, + 367, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 367, + 236 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, 33:6840–6851, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 212, + 505, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 241, + 504, + 265 + ], + "lines": [ + { + "bbox": [ + 105, + 242, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 242, + 505, + 255 + ], + "score": 1.0, + "content": "Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 253, + 404, + 266 + ], + "spans": [ + { + "bbox": [ + 116, + 253, + 404, + 266 + ], + "score": 1.0, + "content": "Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 242, + 505, + 266 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 504, + 295 + ], + "lines": [ + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 285 + ], + "score": 1.0, + "content": "Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 284, + 440, + 295 + ], + "spans": [ + { + "bbox": [ + 116, + 284, + 440, + 295 + ], + "score": 1.0, + "content": "synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1–11, 2016.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 272, + 505, + 295 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 301, + 503, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 314 + ], + "score": 1.0, + "content": "Yanli Ji, Feixiang Xu, Yang Yang, Fumin Shen, Heng Tao Shen, and Wei-Shi Zheng. A large-scale", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 115, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "rgb-d database for arbitrary-view human action recognition. In Proceedings of the 26th ACM", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 324, + 372, + 336 + ], + "spans": [ + { + "bbox": [ + 116, + 324, + 372, + 336 + ], + "score": 1.0, + "content": "international Conference on Multimedia, pp. 1510–1518, 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 302, + 505, + 336 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 342, + 504, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Con-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 116, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "volutional autoencoders for human motion infilling. In 2020 International Conference on 3D", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 117, + 365, + 279, + 377 + ], + "spans": [ + { + "bbox": [ + 117, + 365, + 279, + 377 + ], + "score": 1.0, + "content": "Vision (3DV), pp. 918–927. IEEE, 2020.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 342, + 505, + 377 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 503, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 394, + 322, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 394, + 322, + 407 + ], + "score": 1.0, + "content": "& editing. arXiv preprint arXiv:2209.00349, 2022.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 106, + 383, + 505, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 412, + 503, + 436 + ], + "lines": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 146, + 426 + ], + "score": 1.0, + "content": "Diederik", + "type": "text" + }, + { + "bbox": [ + 146, + 414, + 155, + 424 + ], + "score": 0.29, + "content": "\\mathrm { \\bf P }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 424, + 214, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 214, + 435 + ], + "score": 1.0, + "content": "arXiv:1312.6114, 2013.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 412, + 505, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 504, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 456 + ], + "score": 1.0, + "content": "Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 454, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 116, + 454, + 505, + 467 + ], + "score": 1.0, + "content": "body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 466, + 305, + 477 + ], + "spans": [ + { + "bbox": [ + 116, + 466, + 305, + 477 + ], + "score": 1.0, + "content": "and pattern recognition, pp. 5253–5263, 2020.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 442, + 505, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 504, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 504, + 497 + ], + "score": 1.0, + "content": "Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music con-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 116, + 495, + 272, + 507 + ], + "score": 1.0, + "content": "ditioned 3d dance generation with aist", + "type": "text" + }, + { + "bbox": [ + 273, + 496, + 284, + 505 + ], + "score": 0.38, + "content": "^ { + + }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 495, + 505, + 507 + ], + "score": 1.0, + "content": ". In The IEEE International Conference on Computer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 117, + 506, + 203, + 518 + ], + "spans": [ + { + "bbox": [ + 117, + 506, + 203, + 518 + ], + "score": 1.0, + "content": "Vision (ICCV), 2021.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 483, + 505, + 518 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceed-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 535, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 115, + 535, + 505, + 549 + ], + "score": 1.0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2837–2845,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 545, + 143, + 559 + ], + "spans": [ + { + "bbox": [ + 114, + 545, + 143, + 559 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 524, + 505, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 564, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 116, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "AMASS: Archive of motion capture as surface shapes. In International Conference on Computer", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 117, + 587, + 271, + 599 + ], + "spans": [ + { + "bbox": [ + 117, + 587, + 271, + 599 + ], + "score": 1.0, + "content": "Vision, pp. 5442–5451, October 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 565, + 505, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 504, + 640 + ], + "lines": [ + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 605, + 505, + 619 + ], + "score": 1.0, + "content": "Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recur-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "rent neural networks. In Proceedings of the IEEE conference on computer vision and pattern", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 628, + 257, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 628, + 257, + 640 + ], + "score": 1.0, + "content": "recognition, pp. 2891–2900, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 104, + 605, + 505, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 658, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 116, + 658, + 505, + 670 + ], + "score": 1.0, + "content": "Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 669, + 399, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 399, + 681 + ], + "score": 1.0, + "content": "text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 646, + 505, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 504, + 699 + ], + "score": 1.0, + "content": "Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dim-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "itrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR),", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 721, + 216, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 216, + 732 + ], + "score": 1.0, + "content": "pp. 10975–10985, 2019.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 106, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Mathis Petrovich, Michael J. Black, and Gul Varol. Action-conditioned 3D human motion synthesis ¨", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "with transformer VAE. In International Conference on Computer Vision (ICCV), pp. 10985–", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 105, + 207, + 115 + ], + "spans": [ + { + "bbox": [ + 117, + 105, + 207, + 115 + ], + "score": 1.0, + "content": "10995, October 2021.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "Mathis Petrovich, Michael J. Black, and Gul Varol. TEMOS: Generating diverse human motions ¨", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 133, + 466, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 133, + 466, + 146 + ], + "score": 1.0, + "content": "from textual descriptions. In European Conference on Computer Vision (ECCV), 2022.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 152, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 104, + 150, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 104, + 150, + 506, + 167 + ], + "score": 1.0, + "content": "Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 223, + 175 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 223, + 175 + ], + "score": 1.0, + "content": "data, 4(4):236–252, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 108, + 181, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 504, + 195 + ], + "score": 1.0, + "content": "Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel Cohen-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "Or. Modi: Unconditional motion synthesis from diverse data. arXiv preprint arXiv:2206.08010,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 143, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 143, + 216 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 222, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 116, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "models from natural language supervision. In International Conference on Machine Learning,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 256, + 239, + 267 + ], + "spans": [ + { + "bbox": [ + 114, + 256, + 239, + 267 + ], + "score": 1.0, + "content": "pp. 8748–8763. PMLR, 2021.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 502, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 274, + 504, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 504, + 287 + ], + "score": 1.0, + "content": "Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 285, + 466, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 466, + 298 + ], + "score": 1.0, + "content": "conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 313, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 114, + 313, + 506, + 329 + ], + "score": 1.0, + "content": "networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 115, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 337, + 248, + 349 + ], + "spans": [ + { + "bbox": [ + 116, + 337, + 248, + 349 + ], + "score": 1.0, + "content": "IJCNLP), pp. 3982–3992, 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 355, + 504, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 504, + 367 + ], + "score": 1.0, + "content": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 117, + 366, + 504, + 379 + ], + "spans": [ + { + "bbox": [ + 117, + 366, + 504, + 379 + ], + "score": 1.0, + "content": "cal image segmentation. In International Conference on Medical image computing and computer-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 377, + 322, + 390 + ], + "spans": [ + { + "bbox": [ + 116, + 377, + 322, + 390 + ], + "score": 1.0, + "content": "assisted intervention, pp. 234–241. Springer, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 417, + 310, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 310, + 430 + ], + "score": 1.0, + "content": "2022 Conference Proceedings, pp. 1–10, 2022a.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kam-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "yar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 469, + 224, + 481 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 224, + 481 + ], + "score": 1.0, + "content": "arXiv:2205.11487, 2022b.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "Mingyi Shi, Kfir Aberman, Andreas Aristidou, Taku Komura, Dani Lischinski, Daniel Cohen-Or,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "and Baoquan Chen. Motionet: 3d human motion reconstruction from monocular video with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 510, + 438, + 522 + ], + "spans": [ + { + "bbox": [ + 116, + 510, + 438, + 522 + ], + "score": 1.0, + "content": "skeleton consistency. ACM Transactions on Graphics (TOG), 40(1):1–15, 2020.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 505, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 117, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "learning using nonequilibrium thermodynamics. In International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 551, + 256, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 256, + 561 + ], + "score": 1.0, + "content": "ing, pp. 2256–2265. PMLR, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 104, + 569, + 504, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 580, + 258, + 593 + ], + "spans": [ + { + "bbox": [ + 114, + 580, + 258, + 593 + ], + "score": 1.0, + "content": "preprint arXiv:2010.02502, 2020a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 104, + 598, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "Yang Song and Stefano Ermon. Improved techniques for training score-based generative models.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 609, + 423, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 423, + 622 + ], + "score": 1.0, + "content": "Advances in neural information processing systems, 33:12438–12448, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 650, + 224, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 224, + 662 + ], + "score": 1.0, + "content": "arXiv:2011.13456, 2020b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Ex-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 468, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 468, + 692 + ], + "score": 1.0, + "content": "posing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural informa-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 255, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 255, + 733 + ], + "score": 1.0, + "content": "tion processing systems, 30, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Mathis Petrovich, Michael J. Black, and Gul Varol. Action-conditioned 3D human motion synthesis ¨", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "with transformer VAE. In International Conference on Computer Vision (ICCV), pp. 10985–", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 105, + 207, + 115 + ], + "spans": [ + { + "bbox": [ + 117, + 105, + 207, + 115 + ], + "score": 1.0, + "content": "10995, October 2021.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 136 + ], + "score": 1.0, + "content": "Mathis Petrovich, Michael J. Black, and Gul Varol. TEMOS: Generating diverse human motions ¨", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 133, + 466, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 133, + 466, + 146 + ], + "score": 1.0, + "content": "from textual descriptions. In European Conference on Computer Vision (ECCV), 2022.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 122, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 152, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 104, + 150, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 104, + 150, + 506, + 167 + ], + "score": 1.0, + "content": "Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 223, + 175 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 223, + 175 + ], + "score": 1.0, + "content": "data, 4(4):236–252, 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 150, + 506, + 175 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 181, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 504, + 195 + ], + "score": 1.0, + "content": "Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel Cohen-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 115, + 193, + 505, + 206 + ], + "score": 1.0, + "content": "Or. Modi: Unconditional motion synthesis from diverse data. arXiv preprint arXiv:2206.08010,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 203, + 143, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 143, + 216 + ], + "score": 1.0, + "content": "2022.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 181, + 505, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 222, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 116, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "models from natural language supervision. In International Conference on Machine Learning,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 256, + 239, + 267 + ], + "spans": [ + { + "bbox": [ + 114, + 256, + 239, + 267 + ], + "score": 1.0, + "content": "pp. 8748–8763. PMLR, 2021.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 223, + 506, + 267 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 502, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 274, + 504, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 504, + 287 + ], + "score": 1.0, + "content": "Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 285, + 466, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 466, + 298 + ], + "score": 1.0, + "content": "conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 274, + 504, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 313, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 114, + 313, + 506, + 329 + ], + "score": 1.0, + "content": "networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 115, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 337, + 248, + 349 + ], + "spans": [ + { + "bbox": [ + 116, + 337, + 248, + 349 + ], + "score": 1.0, + "content": "IJCNLP), pp. 3982–3992, 2019.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 303, + 506, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 106, + 355, + 504, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 504, + 367 + ], + "score": 1.0, + "content": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 117, + 366, + 504, + 379 + ], + "spans": [ + { + "bbox": [ + 117, + 366, + 504, + 379 + ], + "score": 1.0, + "content": "cal image segmentation. In International Conference on Medical image computing and computer-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 377, + 322, + 390 + ], + "spans": [ + { + "bbox": [ + 116, + 377, + 322, + 390 + ], + "score": 1.0, + "content": "assisted intervention, pp. 234–241. Springer, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 355, + 504, + 390 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 417, + 310, + 430 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 310, + 430 + ], + "score": 1.0, + "content": "2022 Conference Proceedings, pp. 1–10, 2022a.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 396, + 505, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 449 + ], + "score": 1.0, + "content": "Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kam-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 114, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "yar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 115, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 469, + 224, + 481 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 224, + 481 + ], + "score": 1.0, + "content": "arXiv:2205.11487, 2022b.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 437, + 506, + 481 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 500 + ], + "score": 1.0, + "content": "Mingyi Shi, Kfir Aberman, Andreas Aristidou, Taku Komura, Dani Lischinski, Daniel Cohen-Or,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 116, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "and Baoquan Chen. Motionet: 3d human motion reconstruction from monocular video with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 510, + 438, + 522 + ], + "spans": [ + { + "bbox": [ + 116, + 510, + 438, + 522 + ], + "score": 1.0, + "content": "skeleton consistency. ACM Transactions on Graphics (TOG), 40(1):1–15, 2020.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 488, + 505, + 522 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 505, + 562 + ], + "lines": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 541 + ], + "score": 1.0, + "content": "Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 540, + 504, + 551 + ], + "spans": [ + { + "bbox": [ + 117, + 540, + 504, + 551 + ], + "score": 1.0, + "content": "learning using nonequilibrium thermodynamics. In International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 551, + 256, + 561 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 256, + 561 + ], + "score": 1.0, + "content": "ing, pp. 2256–2265. PMLR, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 527, + 505, + 561 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 569, + 504, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 580, + 258, + 593 + ], + "spans": [ + { + "bbox": [ + 114, + 580, + 258, + 593 + ], + "score": 1.0, + "content": "preprint arXiv:2010.02502, 2020a.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 568, + 506, + 593 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 598, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "Yang Song and Stefano Ermon. Improved techniques for training score-based generative models.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 609, + 423, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 423, + 622 + ], + "score": 1.0, + "content": "Advances in neural information processing systems, 33:12438–12448, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 598, + 505, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 628, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 650, + 224, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 224, + 662 + ], + "score": 1.0, + "content": "arXiv:2011.13456, 2020b.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 628, + 505, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Ex-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 468, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 468, + 692 + ], + "score": 1.0, + "content": "posing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 668, + 505, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural informa-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 255, + 733 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 255, + 733 + ], + "score": 1.0, + "content": "tion processing systems, 30, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 698, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 218, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 218, + 115 + ], + "score": 1.0, + "content": "arXiv:2208.15001, 2022.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 109, + 148, + 272, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 148, + 274, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 274, + 162 + ], + "score": 1.0, + "content": "A ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 108, + 177, + 242, + 189 + ], + "lines": [ + { + "bbox": [ + 106, + 177, + 243, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 243, + 190 + ], + "score": 1.0, + "content": "A.1 DIFFUSION PARAMETERS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 200, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 200, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 505, + 212 + ], + "score": 1.0, + "content": "Learning MDM with different numbers of diffusion steps significantly affects performance and holds", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 448, + 223 + ], + "score": 1.0, + "content": "the potential to accelerate inference time. Table 6 shows optimal performance for", + "type": "text" + }, + { + "bbox": [ + 449, + 211, + 489, + 221 + ], + "score": 0.9, + "content": "T = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 211, + 505, + 223 + ], + "score": 1.0, + "content": ", in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 394, + 235 + ], + "score": 1.0, + "content": "addition, it enables accelerating inference by a factor of 10 compared to", + "type": "text" + }, + { + "bbox": [ + 394, + 222, + 436, + 232 + ], + "score": 0.89, + "content": "T = 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 221, + 505, + 235 + ], + "score": 1.0, + "content": ", which is widely", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 234, + 174, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 174, + 246 + ], + "score": 1.0, + "content": "used for images.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "table", + "bbox": [ + 127, + 261, + 478, + 369 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 127, + 261, + 478, + 369 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 261, + 478, + 369 + ], + "spans": [ + { + "bbox": [ + 127, + 261, + 478, + 369 + ], + "score": 0.98, + "html": "
Diffusion steps (T)R Precision (top 3)↑FID↓Multimodal Dist↓Diversity→
Real0.797±.0020.002±.0002.974±.0089.503±.065
100.574±.0061.461±.0885.816±.0339.369±.058
1000.640±.0070.454±.0395.336±.0299.906±.053
5000.662±.0070.553±.0555.177±.0289.890±.074
10000.611±.0070.544±.0445.566±.0279.559±.086
", + "type": "table", + "image_path": "e2fec044e57faea9d2df7a47f4b8d49941af16ad176262c5e777eeb8e6e950ab.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 127, + 261, + 478, + 297.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 127, + 297.0, + 478, + 333.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 127, + 333.0, + 478, + 369.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 376, + 506, + 410 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "Table 6: Diffusion steps (HumanML3D test set). We run all the evaluation 20 times. Bold indi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 304, + 399 + ], + "score": 1.0, + "content": "cates best result, underline indicates second best,", + "type": "text" + }, + { + "bbox": [ + 304, + 389, + 314, + 399 + ], + "score": 0.7, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 387, + 352, + 399 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 352, + 388, + 372, + 398 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 387, + 454, + 399 + ], + "score": 1.0, + "content": "confidence interval,", + "type": "text" + }, + { + "bbox": [ + 454, + 389, + 467, + 398 + ], + "score": 0.83, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 398, + 214, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 214, + 410 + ], + "score": 1.0, + "content": "that closer to real is better.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 222, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 223, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 223, + 456 + ], + "score": 1.0, + "content": "A.2 GEOMETRIC LOSSES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "We conduct a thorough experiment to evaluate the contribution of geometric losses with the Human-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "score": 1.0, + "content": "Act12 dataset. The results are presented in Table 7. For alignment with prior work, all metrics are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "score": 1.0, + "content": "calculated using the deep features of the action recognition network suggested by Guo et al. (2020).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 500, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 511 + ], + "score": 1.0, + "content": "In general, MDM scores are too close to the real test distribution (i.e. the evaluation network fails to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "discriminate between the two). This means that quantitative results comparing the different variants", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "MDM are too similar to evaluate. As a result, we are not able to decide what combination of geo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "metric losses is preferred. We leave for future work experimenting with a different, more expressive,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 544, + 187, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 187, + 555 + ], + "score": 1.0, + "content": "evaluation network.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "table", + "bbox": [ + 120, + 568, + 493, + 686 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 120, + 568, + 493, + 686 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 120, + 568, + 493, + 686 + ], + "spans": [ + { + "bbox": [ + 120, + 568, + 493, + 686 + ], + "score": 0.984, + "html": "
MethodFID↓Accuracy↑Diversity→Multimodality→
Real0.050±.0000.990±.0006.880±.0202.590±.010
MDM (ours)0.100±.0000.990±.0006.860±.0502.520±.010
w/o foot contact0.080±.0000.990±.0006.810±.0102.580±.010
w/o geometric losses0.090±.0000.990±.0006.820±.0202.550±.010
foot contact only0.100±.0000.990±.0006.860±.0502.520±.010
velocity only0.100±.0000.990±.0006.820±.0202.590±.000
pose only0.090±.0000.990±.0006.830±.0202.570±.020
", + "type": "table", + "image_path": "328d4ad3f45e6215f84fbbaa5b0be5c491d036657b33b272e8c9d01cf9d28fc0.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 120, + 568, + 493, + 607.3333333333334 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 120, + 607.3333333333334, + 493, + 646.6666666666667 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 120, + 646.6666666666667, + 493, + 686.0000000000001 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 694, + 506, + 717 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 693, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 422, + 707 + ], + "score": 1.0, + "content": "Table 7: Geometric losses ablation study. (HumanAct12 dataset) the relative", + "type": "text" + }, + { + "bbox": [ + 422, + 695, + 429, + 705 + ], + "score": 0.53, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 693, + 505, + 707 + ], + "score": 1.0, + "content": "equals 1 when the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 705, + 299, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 299, + 717 + ], + "score": 1.0, + "content": "loss term is included, and 0 when it is excluded.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + } + ], + "index": 26.25 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 218, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 218, + 115 + ], + "score": 1.0, + "content": "arXiv:2208.15001, 2022.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 115 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 148, + 272, + 161 + ], + "lines": [ + { + "bbox": [ + 106, + 148, + 274, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 274, + 162 + ], + "score": 1.0, + "content": "A ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 108, + 177, + 242, + 189 + ], + "lines": [ + { + "bbox": [ + 106, + 177, + 243, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 243, + 190 + ], + "score": 1.0, + "content": "A.1 DIFFUSION PARAMETERS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 200, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 200, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 505, + 212 + ], + "score": 1.0, + "content": "Learning MDM with different numbers of diffusion steps significantly affects performance and holds", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 211, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 211, + 448, + 223 + ], + "score": 1.0, + "content": "the potential to accelerate inference time. Table 6 shows optimal performance for", + "type": "text" + }, + { + "bbox": [ + 449, + 211, + 489, + 221 + ], + "score": 0.9, + "content": "T = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 211, + 505, + 223 + ], + "score": 1.0, + "content": ", in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 394, + 235 + ], + "score": 1.0, + "content": "addition, it enables accelerating inference by a factor of 10 compared to", + "type": "text" + }, + { + "bbox": [ + 394, + 222, + 436, + 232 + ], + "score": 0.89, + "content": "T = 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 221, + 505, + 235 + ], + "score": 1.0, + "content": ", which is widely", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 234, + 174, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 174, + 246 + ], + "score": 1.0, + "content": "used for images.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 200, + 505, + 246 + ] + }, + { + "type": "table", + "bbox": [ + 127, + 261, + 478, + 369 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 127, + 261, + 478, + 369 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 261, + 478, + 369 + ], + "spans": [ + { + "bbox": [ + 127, + 261, + 478, + 369 + ], + "score": 0.98, + "html": "
Diffusion steps (T)R Precision (top 3)↑FID↓Multimodal Dist↓Diversity→
Real0.797±.0020.002±.0002.974±.0089.503±.065
100.574±.0061.461±.0885.816±.0339.369±.058
1000.640±.0070.454±.0395.336±.0299.906±.053
5000.662±.0070.553±.0555.177±.0289.890±.074
10000.611±.0070.544±.0445.566±.0279.559±.086
", + "type": "table", + "image_path": "e2fec044e57faea9d2df7a47f4b8d49941af16ad176262c5e777eeb8e6e950ab.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 127, + 261, + 478, + 297.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 127, + 297.0, + 478, + 333.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 127, + 333.0, + 478, + 369.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 376, + 506, + 410 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "Table 6: Diffusion steps (HumanML3D test set). We run all the evaluation 20 times. Bold indi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 304, + 399 + ], + "score": 1.0, + "content": "cates best result, underline indicates second best,", + "type": "text" + }, + { + "bbox": [ + 304, + 389, + 314, + 399 + ], + "score": 0.7, + "content": "\\pm", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 387, + 352, + 399 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 352, + 388, + 372, + 398 + ], + "score": 0.86, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 387, + 454, + 399 + ], + "score": 1.0, + "content": "confidence interval,", + "type": "text" + }, + { + "bbox": [ + 454, + 389, + 467, + 398 + ], + "score": 0.83, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 398, + 214, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 214, + 410 + ], + "score": 1.0, + "content": "that closer to real is better.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 222, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 223, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 223, + 456 + ], + "score": 1.0, + "content": "A.2 GEOMETRIC LOSSES", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "We conduct a thorough experiment to evaluate the contribution of geometric losses with the Human-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 506, + 489 + ], + "score": 1.0, + "content": "Act12 dataset. The results are presented in Table 7. For alignment with prior work, all metrics are", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 504, + 500 + ], + "score": 1.0, + "content": "calculated using the deep features of the action recognition network suggested by Guo et al. (2020).", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 500, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 511 + ], + "score": 1.0, + "content": "In general, MDM scores are too close to the real test distribution (i.e. the evaluation network fails to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 523 + ], + "score": 1.0, + "content": "discriminate between the two). This means that quantitative results comparing the different variants", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "MDM are too similar to evaluate. As a result, we are not able to decide what combination of geo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "metric losses is preferred. We leave for future work experimenting with a different, more expressive,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 544, + 187, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 187, + 555 + ], + "score": 1.0, + "content": "evaluation network.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 466, + 506, + 555 + ] + }, + { + "type": "table", + "bbox": [ + 120, + 568, + 493, + 686 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 120, + 568, + 493, + 686 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 120, + 568, + 493, + 686 + ], + "spans": [ + { + "bbox": [ + 120, + 568, + 493, + 686 + ], + "score": 0.984, + "html": "
MethodFID↓Accuracy↑Diversity→Multimodality→
Real0.050±.0000.990±.0006.880±.0202.590±.010
MDM (ours)0.100±.0000.990±.0006.860±.0502.520±.010
w/o foot contact0.080±.0000.990±.0006.810±.0102.580±.010
w/o geometric losses0.090±.0000.990±.0006.820±.0202.550±.010
foot contact only0.100±.0000.990±.0006.860±.0502.520±.010
velocity only0.100±.0000.990±.0006.820±.0202.590±.000
pose only0.090±.0000.990±.0006.830±.0202.570±.020
", + "type": "table", + "image_path": "328d4ad3f45e6215f84fbbaa5b0be5c491d036657b33b272e8c9d01cf9d28fc0.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 120, + 568, + 493, + 607.3333333333334 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 120, + 607.3333333333334, + 493, + 646.6666666666667 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 120, + 646.6666666666667, + 493, + 686.0000000000001 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 106, + 694, + 506, + 717 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 693, + 505, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 422, + 707 + ], + "score": 1.0, + "content": "Table 7: Geometric losses ablation study. (HumanAct12 dataset) the relative", + "type": "text" + }, + { + "bbox": [ + 422, + 695, + 429, + 705 + ], + "score": 0.53, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 693, + 505, + 707 + ], + "score": 1.0, + "content": "equals 1 when the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 705, + 299, + 717 + ], + "spans": [ + { + "bbox": [ + 106, + 705, + 299, + 717 + ], + "score": 1.0, + "content": "loss term is included, and 0 when it is excluded.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + } + ], + "index": 26.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 249, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "score": 1.0, + "content": "B EVALUATION METRICS.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 108, + 505, + 142 + ], + "lines": [ + { + "bbox": [ + 105, + 108, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 505, + 120 + ], + "score": 1.0, + "content": "For the completeness of our work, we describe here the quantitative metrics used throughout the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 119, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 506, + 132 + ], + "score": 1.0, + "content": "paper, as they originally described and implemented by Guo et al. (2020) for action-to-motion and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 270, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 270, + 143 + ], + "score": 1.0, + "content": "by Guo et al. (2022a) for text-to-motion.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 108, + 159, + 223, + 170 + ], + "lines": [ + { + "bbox": [ + 105, + 158, + 224, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 224, + 172 + ], + "score": 1.0, + "content": "B.1 ACTION-TO-MOTION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "The following metrics are based on an RNN action recognition network as it was originally trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 192, + 349, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 349, + 204 + ], + "score": 1.0, + "content": "by Guo et al. (2020). We refer to it as the evaluator network.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 504, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Frechet Inception Distance (FID). A widely used metric to evaluate the overall quality for gener-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "ation tasks. FID is calculated upon features extracted from 1,000 generated motion vs ground truth", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "(real) taken from the test set. To adjust this metric to the motion domain, we extract a deep represen-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "tation of the motion with the evaluator network instead of the inception neural network, originally", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 253, + 339, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 339, + 265 + ], + "score": 1.0, + "content": "used for images. A lower value implies better FID results.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 105, + 270, + 504, + 292 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "Accuracy. We classify 1,000 generated motions using the evaluator network, than we calculate the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 280, + 472, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 472, + 294 + ], + "score": 1.0, + "content": "overall recognition accuracy that indicates the correlation of the motion and its action type.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 309 + ], + "score": 1.0, + "content": "Diversity measures the variance of the generated motions across all action categories. We first", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 309, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 299, + 320 + ], + "score": 1.0, + "content": "randomly sample two subsets of the same size", + "type": "text" + }, + { + "bbox": [ + 299, + 309, + 311, + 320 + ], + "score": 0.88, + "content": "S _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 309, + 505, + 320 + ], + "score": 1.0, + "content": "out of a set of all generated motions across all", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 319, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 210, + 333 + ], + "score": 1.0, + "content": "action categories denoted", + "type": "text" + }, + { + "bbox": [ + 210, + 320, + 264, + 332 + ], + "score": 0.94, + "content": "\\{ \\mathbf { v } _ { 1 } , . . . , \\mathbf { v } _ { S _ { d } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 319, + 282, + 333 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 283, + 320, + 340, + 333 + ], + "score": 0.95, + "content": "\\big \\{ \\mathbf { v } _ { 1 } ^ { \\prime \\prime } , . . . , \\mathbf { v } _ { S _ { d } } ^ { \\prime } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 319, + 507, + 333 + ], + "score": 1.0, + "content": ". The diversity of those sets of motions is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 330, + 145, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 145, + 342 + ], + "score": 1.0, + "content": "defied as", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 343, + 378, + 378 + ], + "lines": [ + { + "bbox": [ + 233, + 343, + 378, + 378 + ], + "spans": [ + { + "bbox": [ + 233, + 343, + 378, + 378 + ], + "score": 0.95, + "content": "\\mathrm { D i v e r s i t y } = \\frac { 1 } { S _ { d } } \\sum _ { i = 1 } ^ { S _ { d } } \\parallel \\mathbf { v } _ { i } - \\mathbf { v } _ { i } ^ { \\prime } \\parallel _ { 2 } .", + "type": "interline_equation", + "image_path": "3355b57415a6194344bff9111e3022e8e7a9f043e753343692e778c758f0c6ee.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 343, + 378, + 360.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 233, + 360.5, + 378, + 378.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 138, + 397 + ], + "score": 1.0, + "content": "We use", + "type": "text" + }, + { + "bbox": [ + 139, + 385, + 180, + 396 + ], + "score": 0.91, + "content": "S _ { d } = 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "for our experiments. The diversity value is considered better when closer to the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 395, + 246, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 246, + 408 + ], + "score": 1.0, + "content": "diversity value of the ground truth.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "Multimodality measures the generated motions diversify within each action class. We randomly", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 226, + 438 + ], + "score": 1.0, + "content": "sample two subsets with size", + "type": "text" + }, + { + "bbox": [ + 226, + 424, + 237, + 435 + ], + "score": 0.87, + "content": "S _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 423, + 342, + 438 + ], + "score": 1.0, + "content": "of the same motion class", + "type": "text" + }, + { + "bbox": [ + 342, + 425, + 348, + 434 + ], + "score": 0.26, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 424, + 417, + 436 + ], + "score": 0.89, + "content": "\\{ \\mathbf { v } _ { c , 1 } , \\ldots , \\mathbf { v } _ { c , S _ { l } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 423, + 437, + 438 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 424, + 501, + 437 + ], + "score": 0.93, + "content": "\\{ \\mathbf { v } _ { c , 1 } ^ { \\prime } , . . . , \\mathbf { v } _ { c , S _ { l } } ^ { \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 423, + 506, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 434, + 329, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 263, + 447 + ], + "score": 1.0, + "content": "The multimodality of all action classes", + "type": "text" + }, + { + "bbox": [ + 264, + 435, + 273, + 444 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 434, + 329, + 447 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 201, + 454, + 409, + 489 + ], + "lines": [ + { + "bbox": [ + 201, + 454, + 409, + 489 + ], + "spans": [ + { + "bbox": [ + 201, + 454, + 409, + 489 + ], + "score": 0.93, + "content": "\\mathrm { M u l t i m o d a l i t y } = \\frac { 1 } { C \\times S _ { l } } \\sum _ { c = 1 } ^ { C } \\sum _ { i = 1 } ^ { S _ { l } } \\left\\| \\mathbf { v } _ { c , i } - \\mathbf { v } _ { c , i } ^ { \\prime } \\right\\| _ { 2 } .", + "type": "interline_equation", + "image_path": "572d12d24732b9d2cb80866ed300b6a0e5d20bc1f029326f9d5ffe8d676a6419.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 201, + 454, + 409, + 471.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 201, + 471.5, + 409, + 489.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 497, + 255, + 509 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 256, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 137, + 510 + ], + "score": 1.0, + "content": "We use", + "type": "text" + }, + { + "bbox": [ + 138, + 497, + 172, + 509 + ], + "score": 0.91, + "content": "S _ { l } = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 496, + 256, + 510 + ], + "score": 1.0, + "content": "for our experiments.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 525, + 212, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 213, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 213, + 539 + ], + "score": 1.0, + "content": "B.2 TEXT-TO-MOTION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 548, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 548, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 506, + 561 + ], + "score": 1.0, + "content": "Originally suggested by Guo et al. (2022a), the following metrics are based on a text feature extractor", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "and motion feature extractor jointly trained under contrastive loss to produce geometrically close", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 569, + 353, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 353, + 583 + ], + "score": 1.0, + "content": "feature vectors for matched text-motion pairs, and vise versa.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 587, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "R Precision. (top-3) For each generated motion, its ground-truth text and a randomly selected miss-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 599, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 504, + 609 + ], + "score": 1.0, + "content": "matched descriptions from the test set. We calculate the euclidean distance between the motion", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 622 + ], + "score": 1.0, + "content": "feature and text feature of each description in the pool. We count the average accuracy at top-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "3 places. If the ground truth entry falling into the top-3 candidates, we treat it as True Positive", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 630, + 349, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 349, + 643 + ], + "score": 1.0, + "content": "retrieval. We use a batch size 32 (i.e. 31 negative examples).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 457, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 459, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 459, + 660 + ], + "score": 1.0, + "content": "FID. Same as for action-to-motion, using the motion extractor as the evaluator network.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 664, + 504, + 698 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "Multimodal Distance. We calculate the multimodal distance as the average Euclidean distance", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 675, + 504, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 504, + 688 + ], + "score": 1.0, + "content": "between the motion feature of each generated motion and the text feature of its corresponding de-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 687, + 389, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 389, + 699 + ], + "score": 1.0, + "content": "scription in test set. A lower value implies better multimodal distance.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 106, + 703, + 347, + 715 + ], + "lines": [ + { + "bbox": [ + 106, + 702, + 348, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 304, + 716 + ], + "score": 1.0, + "content": "Diversity. Same as for action-to-motion but with", + "type": "text" + }, + { + "bbox": [ + 304, + 704, + 344, + 715 + ], + "score": 0.89, + "content": "S _ { d } = 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 702, + 348, + 716 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 368, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 369, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 327, + 734 + ], + "score": 1.0, + "content": "Multimodality. Same as for action-to-motion but with", + "type": "text" + }, + { + "bbox": [ + 327, + 721, + 365, + 732 + ], + "score": 0.91, + "content": "S _ { m } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 720, + 369, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 249, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 250, + 96 + ], + "score": 1.0, + "content": "B EVALUATION METRICS.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 108, + 505, + 142 + ], + "lines": [ + { + "bbox": [ + 105, + 108, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 505, + 120 + ], + "score": 1.0, + "content": "For the completeness of our work, we describe here the quantitative metrics used throughout the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 119, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 119, + 506, + 132 + ], + "score": 1.0, + "content": "paper, as they originally described and implemented by Guo et al. (2020) for action-to-motion and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 270, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 270, + 143 + ], + "score": 1.0, + "content": "by Guo et al. (2022a) for text-to-motion.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 108, + 506, + 143 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 159, + 223, + 170 + ], + "lines": [ + { + "bbox": [ + 105, + 158, + 224, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 224, + 172 + ], + "score": 1.0, + "content": "B.1 ACTION-TO-MOTION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 504, + 204 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "The following metrics are based on an RNN action recognition network as it was originally trained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 192, + 349, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 349, + 204 + ], + "score": 1.0, + "content": "by Guo et al. (2020). We refer to it as the evaluator network.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 181, + 505, + 204 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 504, + 264 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Frechet Inception Distance (FID). A widely used metric to evaluate the overall quality for gener-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "ation tasks. FID is calculated upon features extracted from 1,000 generated motion vs ground truth", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "(real) taken from the test set. To adjust this metric to the motion domain, we extract a deep represen-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "tation of the motion with the evaluator network instead of the inception neural network, originally", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 253, + 339, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 339, + 265 + ], + "score": 1.0, + "content": "used for images. A lower value implies better FID results.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 209, + 505, + 265 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 270, + 504, + 292 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "Accuracy. We classify 1,000 generated motions using the evaluator network, than we calculate the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 280, + 472, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 472, + 294 + ], + "score": 1.0, + "content": "overall recognition accuracy that indicates the correlation of the motion and its action type.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 269, + 505, + 294 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 298, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 309 + ], + "score": 1.0, + "content": "Diversity measures the variance of the generated motions across all action categories. We first", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 309, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 299, + 320 + ], + "score": 1.0, + "content": "randomly sample two subsets of the same size", + "type": "text" + }, + { + "bbox": [ + 299, + 309, + 311, + 320 + ], + "score": 0.88, + "content": "S _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 309, + 505, + 320 + ], + "score": 1.0, + "content": "out of a set of all generated motions across all", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 319, + 507, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 210, + 333 + ], + "score": 1.0, + "content": "action categories denoted", + "type": "text" + }, + { + "bbox": [ + 210, + 320, + 264, + 332 + ], + "score": 0.94, + "content": "\\{ \\mathbf { v } _ { 1 } , . . . , \\mathbf { v } _ { S _ { d } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 319, + 282, + 333 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 283, + 320, + 340, + 333 + ], + "score": 0.95, + "content": "\\big \\{ \\mathbf { v } _ { 1 } ^ { \\prime \\prime } , . . . , \\mathbf { v } _ { S _ { d } } ^ { \\prime } \\big \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 319, + 507, + 333 + ], + "score": 1.0, + "content": ". The diversity of those sets of motions is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 330, + 145, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 145, + 342 + ], + "score": 1.0, + "content": "defied as", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 298, + 507, + 342 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 343, + 378, + 378 + ], + "lines": [ + { + "bbox": [ + 233, + 343, + 378, + 378 + ], + "spans": [ + { + "bbox": [ + 233, + 343, + 378, + 378 + ], + "score": 0.95, + "content": "\\mathrm { D i v e r s i t y } = \\frac { 1 } { S _ { d } } \\sum _ { i = 1 } ^ { S _ { d } } \\parallel \\mathbf { v } _ { i } - \\mathbf { v } _ { i } ^ { \\prime } \\parallel _ { 2 } .", + "type": "interline_equation", + "image_path": "3355b57415a6194344bff9111e3022e8e7a9f043e753343692e778c758f0c6ee.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 343, + 378, + 360.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 233, + 360.5, + 378, + 378.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 138, + 397 + ], + "score": 1.0, + "content": "We use", + "type": "text" + }, + { + "bbox": [ + 139, + 385, + 180, + 396 + ], + "score": 0.91, + "content": "S _ { d } = 2 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "for our experiments. The diversity value is considered better when closer to the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 395, + 246, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 246, + 408 + ], + "score": 1.0, + "content": "diversity value of the ground truth.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 384, + 505, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 412, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "Multimodality measures the generated motions diversify within each action class. We randomly", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 226, + 438 + ], + "score": 1.0, + "content": "sample two subsets with size", + "type": "text" + }, + { + "bbox": [ + 226, + 424, + 237, + 435 + ], + "score": 0.87, + "content": "S _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 423, + 342, + 438 + ], + "score": 1.0, + "content": "of the same motion class", + "type": "text" + }, + { + "bbox": [ + 342, + 425, + 348, + 434 + ], + "score": 0.26, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 424, + 417, + 436 + ], + "score": 0.89, + "content": "\\{ \\mathbf { v } _ { c , 1 } , \\ldots , \\mathbf { v } _ { c , S _ { l } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 423, + 437, + 438 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 424, + 501, + 437 + ], + "score": 0.93, + "content": "\\{ \\mathbf { v } _ { c , 1 } ^ { \\prime } , . . . , \\mathbf { v } _ { c , S _ { l } } ^ { \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 423, + 506, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 434, + 329, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 263, + 447 + ], + "score": 1.0, + "content": "The multimodality of all action classes", + "type": "text" + }, + { + "bbox": [ + 264, + 435, + 273, + 444 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 434, + 329, + 447 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 411, + 506, + 447 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 201, + 454, + 409, + 489 + ], + "lines": [ + { + "bbox": [ + 201, + 454, + 409, + 489 + ], + "spans": [ + { + "bbox": [ + 201, + 454, + 409, + 489 + ], + "score": 0.93, + "content": "\\mathrm { M u l t i m o d a l i t y } = \\frac { 1 } { C \\times S _ { l } } \\sum _ { c = 1 } ^ { C } \\sum _ { i = 1 } ^ { S _ { l } } \\left\\| \\mathbf { v } _ { c , i } - \\mathbf { v } _ { c , i } ^ { \\prime } \\right\\| _ { 2 } .", + "type": "interline_equation", + "image_path": "572d12d24732b9d2cb80866ed300b6a0e5d20bc1f029326f9d5ffe8d676a6419.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 201, + 454, + 409, + 471.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 201, + 471.5, + 409, + 489.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 497, + 255, + 509 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 256, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 137, + 510 + ], + "score": 1.0, + "content": "We use", + "type": "text" + }, + { + "bbox": [ + 138, + 497, + 172, + 509 + ], + "score": 0.91, + "content": "S _ { l } = 2 0", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 496, + 256, + 510 + ], + "score": 1.0, + "content": "for our experiments.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 496, + 256, + 510 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 525, + 212, + 537 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 213, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 213, + 539 + ], + "score": 1.0, + "content": "B.2 TEXT-TO-MOTION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 548, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 548, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 506, + 561 + ], + "score": 1.0, + "content": "Originally suggested by Guo et al. (2022a), the following metrics are based on a text feature extractor", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "and motion feature extractor jointly trained under contrastive loss to produce geometrically close", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 569, + 353, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 353, + 583 + ], + "score": 1.0, + "content": "feature vectors for matched text-motion pairs, and vise versa.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 548, + 506, + 583 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 587, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "R Precision. (top-3) For each generated motion, its ground-truth text and a randomly selected miss-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 599, + 504, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 504, + 609 + ], + "score": 1.0, + "content": "matched descriptions from the test set. We calculate the euclidean distance between the motion", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 505, + 622 + ], + "score": 1.0, + "content": "feature and text feature of each description in the pool. We count the average accuracy at top-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 505, + 632 + ], + "score": 1.0, + "content": "3 places. If the ground truth entry falling into the top-3 candidates, we treat it as True Positive", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 630, + 349, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 349, + 643 + ], + "score": 1.0, + "content": "retrieval. We use a batch size 32 (i.e. 31 negative examples).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 587, + 505, + 643 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 648, + 457, + 659 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 459, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 459, + 660 + ], + "score": 1.0, + "content": "FID. Same as for action-to-motion, using the motion extractor as the evaluator network.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 647, + 459, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 664, + 504, + 698 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "Multimodal Distance. We calculate the multimodal distance as the average Euclidean distance", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 675, + 504, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 504, + 688 + ], + "score": 1.0, + "content": "between the motion feature of each generated motion and the text feature of its corresponding de-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 687, + 389, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 389, + 699 + ], + "score": 1.0, + "content": "scription in test set. A lower value implies better multimodal distance.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 663, + 505, + 699 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 703, + 347, + 715 + ], + "lines": [ + { + "bbox": [ + 106, + 702, + 348, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 304, + 716 + ], + "score": 1.0, + "content": "Diversity. Same as for action-to-motion but with", + "type": "text" + }, + { + "bbox": [ + 304, + 704, + 344, + 715 + ], + "score": 0.89, + "content": "S _ { d } = 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 702, + 348, + 716 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 702, + 348, + 716 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 368, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 720, + 369, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 327, + 734 + ], + "score": 1.0, + "content": "Multimodality. Same as for action-to-motion but with", + "type": "text" + }, + { + "bbox": [ + 327, + 721, + 365, + 732 + ], + "score": 0.91, + "content": "S _ { m } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 720, + 369, + 734 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 106, + 720, + 369, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 269, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 270, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 270, + 96 + ], + "score": 1.0, + "content": "C IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 504, + 128 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 120 + ], + "score": 1.0, + "content": "The full implementation of MDM can be found in our published code2. In addition, the followings", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 384, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 384, + 129 + ], + "score": 1.0, + "content": "are the hyperparameters and model details for all of our experiments.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 133, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 506, + 147 + ], + "score": 1.0, + "content": "Diffusion framework. In all of our experiments, we used an implementation of DDPM (Ho et al.,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 144, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 226, + 159 + ], + "score": 1.0, + "content": "2020) by Dhariwal & Nichol", + "type": "text" + }, + { + "bbox": [ + 226, + 145, + 257, + 156 + ], + "score": 0.64, + "content": "( 2 0 2 1 ) ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 144, + 294, + 159 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 295, + 145, + 342, + 156 + ], + "score": 0.89, + "content": "T = 1 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 144, + 505, + 159 + ], + "score": 1.0, + "content": "diffusion steps, cosine noise scheduling", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 156, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 481, + 169 + ], + "score": 1.0, + "content": "(predefined sigmas). All other hyperparameters are according to the implementation defaults.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Transformer architecture. For our transformer architectures, we used the PyTorch implementa-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 408, + 196 + ], + "score": 1.0, + "content": "tion4. We used 8 transformer layers, 4 attention heads, latent dimension", + "type": "text" + }, + { + "bbox": [ + 409, + 184, + 447, + 194 + ], + "score": 0.9, + "content": "d = 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 183, + 505, + 196 + ], + "score": 1.0, + "content": ", dropout 0.1,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "feed-forward size 1024 and gelu activations. The number of learned parameters for each model is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 207, + 177, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 177, + 217 + ], + "score": 1.0, + "content": "stated in Table 8.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "GRU architecture. We use the PyTorch implementation of GRU (Cho et al., 2014) 5 with two layers", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 494, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 494, + 245 + ], + "score": 1.0, + "content": "and latent dimension 512. The number of learned parameters for each model is stated in Table 8.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table", + "bbox": [ + 207, + 272, + 398, + 352 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 250, + 500, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 249, + 500, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 475, + 264 + ], + "score": 1.0, + "content": "Learning hyperparameters. For all of our experiments, we use batch size 64, learning rate", + "type": "text" + }, + { + "bbox": [ + 475, + 250, + 497, + 261 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 249, + 500, + 264 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 207, + 272, + 398, + 352 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 207, + 272, + 398, + 352 + ], + "spans": [ + { + "bbox": [ + 207, + 272, + 398, + 352 + ], + "score": 0.974, + "html": "
Architecture# Parameters (-106)
Transformer Encoder17.88
TransformerDecoder26.29
+ input token26.29
U-net23.47
GRU4.47
", + "type": "table", + "image_path": "91c4e97cfd198b03973344c786adc67ac1c38f2757b6f589d480cba5d087d2cc.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 272, + 398, + 285.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 207, + 285.3333333333333, + 398, + 298.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 207, + 298.66666666666663, + 398, + 311.99999999999994 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 207, + 311.99999999999994, + 398, + 325.33333333333326 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 207, + 325.33333333333326, + 398, + 338.6666666666666 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 207, + 338.6666666666666, + 398, + 351.9999999999999 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 13.75 + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 505, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "Table 8: The number of learned parameters per architecture for the text-to-motion task. For the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "action-to-motion task, there are additional 512 parameters per-class for the class embeddings mod-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 126, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 126, + 394 + ], + "score": 1.0, + "content": "ule.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 107, + 415, + 196, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 198, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 198, + 432 + ], + "score": 1.0, + "content": "D USER STUDY", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 454 + ], + "score": 1.0, + "content": "In Section 4.1 we conduct a user study for the text-to-motion task. We asked 31 users to choose", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "between MDM and state-of-the-art works in a side-by-side view, with both samples generated from", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "the same text prompt randomly sampled from the KIT test set. We repeated this process with 10", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "samples per model and 10 repetitions per sample. This user study enabled a comparison with the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 485, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 504, + 497 + ], + "score": 1.0, + "content": "recent TEMOS model (Petrovich et al., 2022), which was not included in the HumanML3D bench-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "mark. Fig. 4 shows that most of the time, MDM was preferred over the compared models, and even", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 265, + 520 + ], + "score": 1.0, + "content": "preferred over ground truth samples in", + "type": "text" + }, + { + "bbox": [ + 266, + 507, + 293, + 518 + ], + "score": 0.89, + "content": "4 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "of the cases. This user study was designed to mea-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "sure the precision of the models, i.e. which one better fits the input text. The exact phrasing of the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "question was “Which animation better fits the following description?”. A sample question from this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 219, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 219, + 553 + ], + "score": 1.0, + "content": "study is presented in Fig. 5.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 688, + 343, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 687, + 316, + 700 + ], + "spans": [ + { + "bbox": [ + 118, + 687, + 316, + 700 + ], + "score": 1.0, + "content": "2https://github.com/GuyTevet/motion-diffusion-model", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 698, + 279, + 711 + ], + "spans": [ + { + "bbox": [ + 119, + 698, + 279, + 711 + ], + "score": 1.0, + "content": "3https://github.com/openai/guided-diffusion", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 709, + 264, + 723 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 264, + 723 + ], + "score": 1.0, + "content": "4https://pytorch.org/docs/stable/nn.html", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 720, + 342, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 342, + 733 + ], + "score": 1.0, + "content": "5https://pytorch.org/docs/stable/generated/torch.nn.GRU.html", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 269, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 270, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 270, + 96 + ], + "score": 1.0, + "content": "C IMPLEMENTATION DETAILS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 106, + 504, + 128 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 120 + ], + "score": 1.0, + "content": "The full implementation of MDM can be found in our published code2. In addition, the followings", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 117, + 384, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 117, + 384, + 129 + ], + "score": 1.0, + "content": "are the hyperparameters and model details for all of our experiments.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 104, + 505, + 129 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 133, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 506, + 147 + ], + "score": 1.0, + "content": "Diffusion framework. In all of our experiments, we used an implementation of DDPM (Ho et al.,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 144, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 226, + 159 + ], + "score": 1.0, + "content": "2020) by Dhariwal & Nichol", + "type": "text" + }, + { + "bbox": [ + 226, + 145, + 257, + 156 + ], + "score": 0.64, + "content": "( 2 0 2 1 ) ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 144, + 294, + 159 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 295, + 145, + 342, + 156 + ], + "score": 0.89, + "content": "T = 1 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 144, + 505, + 159 + ], + "score": 1.0, + "content": "diffusion steps, cosine noise scheduling", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 156, + 481, + 169 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 481, + 169 + ], + "score": 1.0, + "content": "(predefined sigmas). All other hyperparameters are according to the implementation defaults.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 133, + 506, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Transformer architecture. For our transformer architectures, we used the PyTorch implementa-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 408, + 196 + ], + "score": 1.0, + "content": "tion4. We used 8 transformer layers, 4 attention heads, latent dimension", + "type": "text" + }, + { + "bbox": [ + 409, + 184, + 447, + 194 + ], + "score": 0.9, + "content": "d = 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 183, + 505, + 196 + ], + "score": 1.0, + "content": ", dropout 0.1,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 506, + 207 + ], + "score": 1.0, + "content": "feed-forward size 1024 and gelu activations. The number of learned parameters for each model is", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 207, + 177, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 177, + 217 + ], + "score": 1.0, + "content": "stated in Table 8.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 172, + 506, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "GRU architecture. We use the PyTorch implementation of GRU (Cho et al., 2014) 5 with two layers", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 494, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 494, + 245 + ], + "score": 1.0, + "content": "and latent dimension 512. The number of learned parameters for each model is stated in Table 8.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 222, + 505, + 245 + ] + }, + { + "type": "table", + "bbox": [ + 207, + 272, + 398, + 352 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 250, + 500, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 249, + 500, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 475, + 264 + ], + "score": 1.0, + "content": "Learning hyperparameters. For all of our experiments, we use batch size 64, learning rate", + "type": "text" + }, + { + "bbox": [ + 475, + 250, + 497, + 261 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 497, + 249, + 500, + 264 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 207, + 272, + 398, + 352 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 207, + 272, + 398, + 352 + ], + "spans": [ + { + "bbox": [ + 207, + 272, + 398, + 352 + ], + "score": 0.974, + "html": "
Architecture# Parameters (-106)
Transformer Encoder17.88
TransformerDecoder26.29
+ input token26.29
U-net23.47
GRU4.47
", + "type": "table", + "image_path": "91c4e97cfd198b03973344c786adc67ac1c38f2757b6f589d480cba5d087d2cc.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 272, + 398, + 285.3333333333333 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 207, + 285.3333333333333, + 398, + 298.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 207, + 298.66666666666663, + 398, + 311.99999999999994 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 207, + 311.99999999999994, + 398, + 325.33333333333326 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 207, + 325.33333333333326, + 398, + 338.6666666666666 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 207, + 338.6666666666666, + 398, + 351.9999999999999 + ], + "spans": [], + "index": 18 + } + ] + } + ], + "index": 13.75 + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 505, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 371 + ], + "score": 1.0, + "content": "Table 8: The number of learned parameters per architecture for the text-to-motion task. For the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "action-to-motion task, there are additional 512 parameters per-class for the class embeddings mod-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 126, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 126, + 394 + ], + "score": 1.0, + "content": "ule.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 359, + 505, + 394 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 415, + 196, + 429 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 198, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 198, + 432 + ], + "score": 1.0, + "content": "D USER STUDY", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 506, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 454 + ], + "score": 1.0, + "content": "In Section 4.1 we conduct a user study for the text-to-motion task. We asked 31 users to choose", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "between MDM and state-of-the-art works in a side-by-side view, with both samples generated from", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 506, + 476 + ], + "score": 1.0, + "content": "the same text prompt randomly sampled from the KIT test set. We repeated this process with 10", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "samples per model and 10 repetitions per sample. This user study enabled a comparison with the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 485, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 504, + 497 + ], + "score": 1.0, + "content": "recent TEMOS model (Petrovich et al., 2022), which was not included in the HumanML3D bench-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "mark. Fig. 4 shows that most of the time, MDM was preferred over the compared models, and even", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 265, + 520 + ], + "score": 1.0, + "content": "preferred over ground truth samples in", + "type": "text" + }, + { + "bbox": [ + 266, + 507, + 293, + 518 + ], + "score": 0.89, + "content": "4 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "of the cases. This user study was designed to mea-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "sure the precision of the models, i.e. which one better fits the input text. The exact phrasing of the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "question was “Which animation better fits the following description?”. A sample question from this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 540, + 219, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 219, + 553 + ], + "score": 1.0, + "content": "study is presented in Fig. 5.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 440, + 506, + 553 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 303, + 443, + 489 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 303, + 443, + 489 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 303, + 443, + 489 + ], + "spans": [ + { + "bbox": [ + 165, + 303, + 443, + 489 + ], + "score": 0.968, + "type": "image", + "image_path": "08840b6b387e4d6c1ee66ffc62da485b77bf28066ea8a312462a7e54006beead.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 165, + 303, + 443, + 365.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 165, + 365.0, + 443, + 427.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 165, + 427.0, + 443, + 489.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 493, + 503, + 506 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "score": 1.0, + "content": "Figure 5: An example question for our text-to-motion user study, using the Google Forms platform.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2023", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 165, + 303, + 443, + 489 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 165, + 303, + 443, + 489 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 303, + 443, + 489 + ], + "spans": [ + { + "bbox": [ + 165, + 303, + 443, + 489 + ], + "score": 0.968, + "type": "image", + "image_path": "08840b6b387e4d6c1ee66ffc62da485b77bf28066ea8a312462a7e54006beead.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 165, + 303, + 443, + 365.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 165, + 365.0, + 443, + 427.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 165, + 427.0, + 443, + 489.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 493, + 503, + 506 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 505, + 507 + ], + "score": 1.0, + "content": "Figure 5: An example question for our text-to-motion user study, using the Google Forms platform.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file