Reproduce “SceneAnything: Agentic Reconstruction of Large-Scale Mesh Scenes from a Single Video” 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
Bowen Xue, Saeed Hadadan, Fabrice Rousselle, Zahra Montazeri, Ravi Ramamoorthi, Milos Hasan
SceneAnything builds editable 3D scenes from video using shared camera, depth, and object-mask estimates, coordinated agents, and revision checks. Its level-of-detail rule improves governed-pixel LPIPS on a garden bed from 0.70 to 0.58. On 10 aligned scenes, novel-view DINOv3 similarity is 0.53 versus at most 0.38, measuring content similarity, not geometric accuracy; unseen geometry remains inferred.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.