SpatialForge-10M

SpatialForge: Bootstrapping 3D-Aware Spatial Reasoning from Open-World 2D Images

📑 Paper

Zishan Liu, Ruoxi Zang, Yanglin Zhang, Wei Liu, Yin Zhang, Jian Yao, Jiayin Zheng, Zhengzhe Liu

Lingnan University · XPENG Robotics

---

📦 SpatialForge-10M

A large-scale vision-language dataset designed for 3D-aware spatial perception and reasoning from open-world 2D images.

SpatialForge-10M contains over 10 million QA pairs generated from 2.8 million curated real-world images, covering both low-level spatial perception and high-level spatial reasoning tasks. The dataset is constructed from diverse open-world image sources including Objects365, Pixmo and OpenImages, enabling broad visual diversity across indoor, outdoor, egocentric, and internet-scale scenes.

SpatialForge-10M is designed for:

• Spatial reasoning pretraining
• Multitask VLM supervision
• 3D-aware perception learning
• Grounding and referring research
• Camera-centric and human-centric reasoning
• Spatial instruction tuning

The dataset supports a unified QA format suitable for training modern multimodal large language models such as Qwen-VL, InternVL, LLaVA, and related architectures.

---

📌 Important Notice

This release contains the full SpatialForge-10M annotations, including all question-answer pairs and task splits. You will need to obtain the corresponding images directly from the source dataset.

Key Features

• ✅ 10M+ spatial QA pairs spanning perception and reasoning tasks
• ✅ 2.8M curated open-world images from diverse visual domains
• ✅ Covers both object-centric and human-centric spatial reasoning
• ✅ Includes grounding, referring, counting, depth reasoning, and perspective understanding
• ✅ Designed for scalable VLM pretraining and instruction tuning
• ✅ Bounding boxes are normalized to [0, 1000], following the format used in Qwen3-VL pretraining

---

🧠 Task Overview

SpatialForge-10M contains six major spatial tasks divided into two categories: Perception and Relation Reasoning.

Level	Task	Description	Count
Perception	Grounding	Localize objects from textual descriptions → bbox prediction	3.6M
Perception	Referring	Generate object descriptions from regions/bboxes	3.6M
Perception	Counting	Count objects satisfying semantic conditions	495k
Relation	Near-Far	Determine relative depth comapring between objects	2.6M
Relation	Left-Right	Infer camera-centric horizontal spatial relations	93k
Relation	Perspective	Human-centric viewpoint and perspective reasoning	8k
	Total		10.2M

---

🌍 Open-World Data Sources

SpatialForge-10M is bootstrapped from large-scale public image datasets:

• Objects365
• OpenImages
• Pixmo

These datasets provide rich scene diversity across:

• Indoor environments
• Outdoor scenes
• Human-object interactions
• Crowded object layouts
• Real-world internet imagery

Our pipeline automatically constructs spatial supervision signals from 2D images while preserving geometric consistency and viewpoint awareness.

---

@article{liu2026spatialforge,
  title={SpatialForge: Bootstrapping 3D-Aware Spatial Reasoning from Open-World 2D Images},
  author={Liu, Zishan and Zang, Ruoxi and Zhang, Yanglin and Liu, Wei and Zhang, Yin and Yao, Jian and Zheng, Jiayin and Liu, Zhengzhe},
  journal={arXiv preprint arXiv:2605.11462},
  year={2026}
}