ABO-I2V

Image-to-video product retrieval derived from Amazon Berkeley Objects (ABO). Built for MTEB as ABOI2VRetrieval.

Each product contributes one video (its 360-degree turntable "spin") and one query image (an independently shot catalog photograph of the same product, in a room or scene). The two assets were captured at different times, in different places, under different lighting, so a query is never a frame of its own positive video and frame leakage is structurally impossible.

config	split	rows	contents
corpus	test	2857	_id, video (h264, 384px long side, 24 frames, 12 fps, 2.0 s)
queries	test	2857	_id, image (catalog photograph, 256px long side)
default	test	2857	query-id, corpus-id, score (1:1, score 1)

Corpus frames are selected by azimuth % 3 == 0, which yields exactly 24 frames at 15-degree steps for every spin in ABO. Product types are restricted to five volumetric home-goods categories (CHAIR, SOFA, TABLE, HOME_FURNITURE_AND_DECOR, LAMP); flat goods such as rugs and wall art are excluded because a turntable rotation of a flat object is close to degenerate. Query images that share a perceptual hash with another product, that fail a zero-shot category gate, that fall within a perceptual-hash radius of the product's own spin, or that are studio shots on a white sweep are all excluded.

License and attribution

This derived dataset is distributed under the Creative Commons Attribution 4.0 International Public License (CC BY 4.0), the license of the source data: https://creativecommons.org/licenses/by/4.0/

Credit for the data, including all images, must be given to:

Amazon.com

Credit for building the source dataset, archives and benchmark sets must be given to:

Matthieu Guillaumin (Amazon.com), Thomas Dideriksen (Amazon.com), Kenan Deng (Amazon.com), Himanshu Arora (Amazon.com), Arnab Dhua (Amazon.com), Xi (Brian) Zhang (Amazon.com), Tomas Yago-Vicente (Amazon.com), Jasmine Collins (UC Berkeley), Shubham Goel (UC Berkeley) and Jitendra Malik (UC Berkeley)

Changes made to the licensed material: turntable spin frame sequences were subsampled to 24 frames and re-encoded to H.264 video at 384px on the long side; catalog images were filtered and subset-selected, and are redistributed at the 256px resolution published in ABO's images/small/. No source pixels were otherwise altered.

Source: https://amazon-berkeley-objects.s3.us-east-1.amazonaws.com/index.html · License text: LICENSE-CC-BY-4.0.txt in the ABO bucket · The licensed material is provided as-is, without warranties or conditions of any kind; see Section 5 of CC BY 4.0 for the full disclaimer of warranties and limitation of liability.

Note on the source license. The ABO bucket still contains a stale LICENSE-CC-BY-NC-4.0.txt dated 2021-06-14, and the AWS Open Data registry entry was never updated after the mid-2023 relicense. The current authoritative terms are CC BY 4.0: LICENSE-CC-BY-4.0.txt (2023-06-26), the bucket README.md (2023-07-18), the spins/, images/, listings/ and 3dmodels/ READMEs, and the project page, which all state CC BY 4.0 and reference only the CC BY 4.0 license file. The CVPR 2022 paper predates the relicense.

Citation

@article{collins2022abo,
  author = {Collins, Jasmine and Goel, Shubham and Deng, Kenan and Luthra, Achleshwar and Xu, Leon and Gundogdu, Erhan and Zhang, Xi and Yago Vicente, Tomas F and Dideriksen, Thomas and Arora, Himanshu and Guillaumin, Matthieu and Malik, Jitendra},
  journal = {CVPR},
  title = {ABO: Dataset and Benchmarks for Real-World 3D Object Understanding},
  year = {2022},
}