Reproduce “OneCanvas: 3D Scene Understanding via Panoramic Reprojection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Bartłomiej Baranowski, Dave Zhenyu Chen, Matthias Nießner

OneCanvas gives vision-language models a shared spatial representation for reasoning across camera views. It backprojects image patches using depth and camera pose, places them on a viewpoint-centered panorama, adds metric positions, and uses procedural geometry pretraining before scene-question adaptation. It reports 65.3 EM@1 on SQA3D, 71.3 on VSI-Bench, and 72.1 on SPBench without its training data. The method still requires depth and camera poses.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.