Reproduce “Seeing the World and the Self from Egocentric 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
Kai Guan, Minchao Jiang, Ruichen WangLi, Wentao Zhu, Lei Zhang
RESELF is a unified framework from The Hong Kong Polytechnic University and Eastern Institute of Technology that jointly recovers 3D scene geometry and the wearer's full-body motion from monocular egocentric video within a shared metric frame. It improves camera trajectory estimation, reducing Absolute Trajectory Error to 0.012m, and lowers human motion reconstruction error, achieving a Mean Per-Joint Position Error of 109.8mm.
Key insights
- Unifying scene and self-reconstruction benefits from a shared metric frame, allowing deterministic geometric regression for the scene and generative motion inference for the occluded body to inform each other.
- Adapting exocentric geometry foundation models to egocentric video requires specific consistency objectives to mitigate challenges like rapid camera motion and dynamic foregrounds.
- Reciprocal information flow, where scene geometry conditions motion generation and reconstructed kinematics refine camera estimation, leads to mutually consistent gains across both tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.