Reproduce “WorldSplat: Gaussian-Centric Feed-Forward 4D Scene Generation 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
Ziyue Zhu, Zhanqian Wu, Zhenxin Zhu, Lijun Zhou, Haiyang Sun, Bing Wan, Kun Ma, Guang Chen and 3 more
WorldSplat, developed by Xiaomi EV and academic partners, introduces a framework for generating dynamic 4D driving scenes. It unifies generative capabilities with 3D consistency by producing an explicit 4D Gaussian representation, enabling high-fidelity multi-track videos and novel views that are spatio-temporally consistent.
Key insights
- Directly generating an explicit 4D Gaussian representation from multi-modal latents allows for both generative flexibility and geometrically consistent novel-view synthesis.
- A feed-forward architecture for 4D scene generation enables scalability and avoids the computational cost of per-scene optimization common in reconstruction methods.
- Decomposition into static and dynamic Gaussians, aggregated in a unified ego-coordinate system, is crucial for maintaining spatio-temporal coherence across multi-frame scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.