Reproduce “RiGS: Rigid-aware 4D Gaussian Splatting from a Single 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
Chenyu Wu, Wanhua Li, Zhu-Tian Chen, Hanspeter Pfister
RiGS reconstructs dynamic 3D scenes from single monocular videos by combining static, rigid, and transient Gaussian primitives within a 4D Gaussian Splatting framework. This approach effectively captures both long-term rigid motions and short-term complex deformations, achieving an average PSNR of 27.43 and LPIPS of 0.051 on the Nvidia Dynamic Scenes Dataset.
Key insights
- Dynamic scenes require a multi-primitive representation to comprehensively model static components, long-term rigid motion, and short-term transient deformations simultaneously.
- Robust static-dynamic decomposition at the object level is crucial for accurate initialization and performance of dynamic Gaussian models, outperforming pixel-level masking by preventing spurious responses.
- Dense 3D scene flow provides essential, geometrically consistent supervision for learning diverse motion patterns, especially for flexible transient primitives, improving motion accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.