Reproduce “Nighttime Autonomous Driving Scene Reconstruction with Physically-Based Gaussian Splatting” 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

Tae-Kyeong Kim, Xingxin Chen, Guile Wu, Chengjie Huang, Dongfeng Bai, Bingbing Liu

Researchers from Huawei Noah's Ark Lab and the University of Toronto developed a framework that integrates physically-based rendering into 3D Gaussian Splatting for photorealistic reconstruction of dynamic nighttime autonomous driving scenes. This approach achieves superior qualitative and quantitative performance on Waymo Open and nuScenes datasets, particularly in rendering complex lighting interactions.

Key insights

  • Explicitly decomposing nighttime illumination into global diffuse (SH-based) and localized specular (ASG-based) components, constrained by a physically-based BRDF model, is essential for robust and photorealistic rendering in low-light dynamic scenes.
  • Anisotropic Spherical Gaussians (ASGs) are more effective than Spherical Harmonics (SH) for capturing sharp, high-frequency specular highlights characteristic of nighttime environments.
  • Integrating normal supervision during optimization significantly improves the geometric accuracy of Gaussian primitives, which is critical for accurate light interaction and surface appearance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.