MuDG: Taming Multi-modal Diffusion with Gaussian Splatting for Urban Scene Reconstruction
This work addresses reconstruction and synthesis challenges in autonomous driving scenes, offering a feed-forward approach that improves robustness under extreme viewpoint changes, though it appears incremental as it builds on existing diffusion and Gaussian splatting techniques.
The paper tackles the problem of 3D urban scene reconstruction and novel view synthesis in autonomous driving, which suffers from performance drops under viewpoint changes and lack of temporal coherence, by proposing MuDG, a framework that integrates multi-modal diffusion with Gaussian splatting to synthesize photorealistic outputs for novel viewpoints, achieving superior performance on the Open Waymo Dataset.
Recent breakthroughs in radiance fields have significantly advanced 3D scene reconstruction and novel view synthesis (NVS) in autonomous driving. Nevertheless, critical limitations persist: reconstruction-based methods exhibit substantial performance deterioration under significant viewpoint deviations from training trajectories, while generation-based techniques struggle with temporal coherence and precise scene controllability. To overcome these challenges, we present MuDG, an innovative framework that integrates Multi-modal Diffusion model with Gaussian Splatting (GS) for Urban Scene Reconstruction. MuDG leverages aggregated LiDAR point clouds with RGB and geometric priors to condition a multi-modal video diffusion model, synthesizing photorealistic RGB, depth, and semantic outputs for novel viewpoints. This synthesis pipeline enables feed-forward NVS without computationally intensive per-scene optimization, providing comprehensive supervision signals to refine 3DGS representations for rendering robustness enhancement under extreme viewpoint changes. Experiments on the Open Waymo Dataset demonstrate that MuDG outperforms existing methods in both reconstruction and synthesis quality.