Reproduce “SplatFlow: Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving” 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
Su Sun, Cheng Zhao, Zhuoyang Sun, Yingjie Victor Chen, Mei Chen
SplatFlow is a self-supervised dynamic Gaussian Splatting framework that reconstructs high-fidelity 4D urban scenes, enabling real-time novel view synthesis and significantly improved rendering of dynamic objects without requiring 3D bounding box annotations. It achieves state-of-the-art performance on Waymo and KITTI datasets, rendering at 40-44 FPS.
Key insights
- Leveraging 3D LiDAR point cloud information to pre-train a Neural Motion Flow Field is highly effective for robust 3D motion estimation and self-supervised static/dynamic scene decomposition in complex urban environments.
- Integrating a continuous 3D motion flow field with a 4D Gaussian representation allows for accurate temporal correspondence modeling of dynamic objects, leading to superior multi-view consistency and reduced rendering artifacts.
- Knowledge distillation from powerful 2D foundation models can effectively enhance the identification and rendering quality of dynamic objects within a 4D scene reconstruction framework, even without explicit 3D supervision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.