Reproduce “MoRGS: Efficient Per-Gaussian Motion Reasoning for Streamable Dynamic 3D Scenes” 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
Wonjoon Lee, Sungmin Woo, Donghyeong Kim, Jungho Lee, Sangheon Park, Sangyoun Lee
MoRGS introduces an efficient online framework for dynamic 3D scene reconstruction using 3D Gaussian Splatting, explicitly modeling per-Gaussian motion with flow-guided supervision and a confidence mechanism. The method achieves state-of-the-art rendering quality and improved temporal consistency among online approaches while maintaining streamable performance.
Key insights
- Explicitly guiding per-Gaussian motion with sparse 2D optical flow, rather than solely relying on photometric loss, improves motion fidelity and reduces ambiguity in dynamic scene reconstruction.
- A learnable 3D motion offset field effectively aggregates multi-view information to refine and stabilize motion estimates, even with limited flow supervision.
- Differentiating between static and dynamic Gaussians through a confidence score allows for targeted attribute updates, preventing spurious motion in static areas and robustly capturing large movements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.