Reproduce “GaussianPrediction: Dynamic 3D Gaussian Prediction for Motion Extrapolation and Free View Synthesis” 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
Boming Zhao, Yuan Li, Ziyu Sun, Lin Zeng, Yujun Shen, Rui Ma, Yinda Zhang, Hujun Bao and 1 more
GaussianPrediction combines dynamic 3D Gaussian Splatting with compact motion forecasting to reconstruct observed scenes, extrapolate their near-future states, and render those predictions from unseen viewpoints. A lifecycle model handles disappearing geometry, while graph-based prediction over a few hundred key points controls more than 200,000 Gaussians; on D-NeRF, the method reaches 24.62 PSNR, 0.939 SSIM, and 0.051 LPIPS for future synthesis.
Key insights
- Modeling proximity in combined spatial-motion space helps distinguish adjacent regions with different dynamics, such as an object being cut and the tool performing the cut.
- Lifecycle-aware opacity prediction allows Gaussians to disappear when surfaces are removed, avoiding persistent residual geometry after irreversible changes.
- Graph-based key-point forecasting improves average future prediction over directly extrapolating the deformation MLP, although the benefit varies across motion sequences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.