Reproduce “Gaussians on their Way: Wasserstein-Constrained 4D Gaussian Splatting with State-Space Modeling” 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
Junli Deng, Yihao Luo
Researchers from Communication University of China and Imperial College London introduce a new framework for 4D Gaussian Splatting that integrates a State Consistency Filter with Wasserstein geometry to model dynamic scenes. The approach significantly improves rendering quality and temporal coherence, achieving a PSNR of 34.45 dB on the D-NeRF dataset and reducing flickering artifacts by providing more physically plausible Gaussian motion.
Key insights
- Treating Gaussian parameters as independent Euclidean entities is insufficient for capturing their true dynamic behavior; a geometric approach like Wasserstein distance inherently accounts for combined positional and shape changes.
- Integrating a state-space modeling filter with optimal transport theory (Wasserstein geometry) provides a principled way to robustly estimate and propagate Gaussian states, reducing noise and artifacts.
- Utilizing logarithmic and exponential maps on the manifold of Gaussian distributions allows for physically plausible and smooth evolution, ensuring predicted covariances remain valid symmetric positive definite matrices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.