Reproduce “MagicRoad: Semantic-Aware 3D Road Surface Reconstruction via Obstacle Inpainting” 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
Xingyue Peng, Yuandong Lyu, Lang Zhang, Jian Zhu, Songtao Wang, Jiaxin Deng, Songxin Lu, Weiliang Ma and 3 more
MagicRoad, developed by Li Auto Inc., achieves robust and geometrically accurate 3D road surface reconstruction from in-car camera videos by employing semantic-aware 2D Gaussian surfels, obstacle inpainting, and color enhancement. The system outperforms state-of-the-art methods like RoGS on the nuScenes dataset, yielding improved visual fidelity and 3D elevation accuracy.
Key insights
- Explicitly modeling and removing dynamic/static occlusions through video inpainting is crucial for reconstructing complete and accurate road surfaces in real-world driving.
- Augmenting a planar-adapted 2D Gaussian surfel representation with semantic features allows for fine-grained, class-aware guidance, improving both geometric stability and appearance consistency.
- Harmonizing appearance variations across multiple views using semantic-aware color enhancement in HSV space leads to more visually coherent and spatially consistent road surface reconstructions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.