Reproduce “World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular 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
Liyuan Zhu, Shengyu Huang, Amrita Mazumdar, Tianye Li, Zan Gojcic, Gordon Wetzstein, Iro Armeni, Shalini De Mello and 1 more
World from Motion refines scenes from monocular video using an initial 3D Gaussian splatting reconstruction as a scaffold. It conditions video generation on rendered geometry, motion, and camera cues, then folds generated views back into the scene. On DyCheck’s static-camera protocol, it improves covisible mPSNR from 19.28 to 20.06 and mLPIPS from 0.246 to 0.187, but depends on a useful initial reconstruction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.