VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations

Accepted to ECCV 2026 🚀

Motivation

Why VIEW2SPACE?
A single glance is often insufficient for completing real-world tasks. Humans naturally integrate observations across sparse viewpoints to form a shared spatial understanding. Given heterogeneous views (e.g., robotic dog and drone), one can efficiently align what different agents observe and reason across sparse viewpoints.

Announcement

VIEW2SPACE resources are publicly available.

• Paper status: Accepted to ECCV 2026 🚀
• Hugging Face collection: VIEW2SPACE
• Testing set release: view2space-v1 ✅
• Training data release: view2space-train ✅
• Checkpoint release: view2space_GCoT_4b_checkpoint ✅
• Scene generation source release: scene_generation_src ✅

Training Data Release

We have released the VIEW2SPACE training data on Hugging Face: view2space-train. The release includes grounded chain-of-thought supervision for multi-view spatial reasoning, connecting sparse visual observations with step-by-step spatial reasoning traces.

Scene Generation

If you want to generate custom 3D scenes for VIEW2SPACE-style data collection, please start from scene_generation_src. The folder contains the Blender-based scene generation pipeline, including asset preprocessing, preview rendering, configurable scene synthesis, and metadata export.

Quick Start

For public evaluation, start here:

• src/README.md
• src/eval/run_all_subsets.sh

Source code is organized by responsibility:

• src/eval: public evaluation entrypoints
• src/train: training launchers and preprocessing helpers
• src: shared prompt, message-building, and config files
• scene_generation_src: Blender-based 3D scene generation pipeline

Citation

If you find VIEW2SPACE useful in your research, please cite:

@article{ke2026view2space,
  title={VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations},
  author={Ke, Fucai and Cai, Zhixi and Li, Boying and Chen, Long and Lin, Beibei and Wang, Weiqing and Haghighi, Pari Delir and Haffari, Gholamreza and Rezatofighi, Hamid},
  journal={arXiv preprint arXiv:2603.16506},
  year={2026}
}