Reproduce “Gaussian-Flow: 4D Reconstruction with Dynamic 3D Gaussian Particle” 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
Youtian Lin, Zuozhuo Dai, Siyu Zhu, Yao Yao
Gaussian-Flow presents an explicit, particle-based deformation model built on 3D Gaussian Splatting for reconstructing and rendering dynamic scenes from video data. The method achieves real-time rendering performance and higher reconstruction quality on dynamic datasets while reducing training time by approximately 5x compared to per-frame 3DGS models.
Key insights
- Combining polynomial fitting in the time domain and Fourier series fitting in the frequency domain within the Dual-Domain Deformation Model effectively captures both smooth and complex motions.
- Introducing adaptive timestamp scaling for each Gaussian allows the model to handle varying motion speeds within a dynamic scene, preventing optimization instability.
- Regularization techniques, including time smoothness loss and KNN rigid loss, are essential for ensuring temporal coherence and maintaining local spatial rigidity during deformations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.