NeurIPS 2025｜智源&清华带来自驾重建新SOTA！

Core Viewpoint - The article discusses a novel multi-scale bilateral grid framework for 3D scene reconstruction in autonomous driving, addressing challenges such as lighting variations and dynamic objects, leading to improved geometric accuracy and visual quality [5][10][39]. Group 1: Methodology - The proposed framework combines the strengths of appearance codes and bilateral grids to achieve efficient and accurate scene reconstruction [11][13]. - The architecture employs Gaussian splatting to model complex driving scenes, decomposing them into a mixed scene graph that includes independent modeling of static and dynamic elements [14]. - The framework consists of three levels: coarse, intermediate, and fine, each addressing different aspects of lighting and detail adjustments [15]. Group 2: Experimental Results - Extensive experiments on datasets like Waymo, NuScenes, Argoverse, and PandaSet demonstrate that the proposed method significantly outperforms existing models in geometric accuracy and appearance consistency [19][39]. - In the Waymo dataset, the chamfer distance (CD) improved from 1.378 (OmniRe) to 0.989, a 28.2% enhancement [21]. - The method achieved a PSNR of 27.69 and an SSIM of 0.847 on the NuScenes dataset, surpassing OmniRe's scores of 26.37 and 0.837 respectively [23]. Group 3: Robustness and Versatility - The framework shows enhanced performance in extreme scenarios such as night scenes and varying lighting conditions, proving its robustness [27][39]. - The method can be integrated as a plug-and-play enhancement module into existing models like ChatSim and StreetGS, resulting in significant improvements in reconstruction quality [25][26]. Group 4: Future Directions - The research team plans to further optimize the framework for larger and more complex scenes and explore more efficient computational methods for practical applications in autonomous driving [40].