Abstract
Realistic scene reconstruction and view synthesis are essential for advancingautonomous driving systems by simulating safety-critical scenarios. 3D GaussianSplatting excels in real-time rendering and static scene reconstructions butstruggles with modeling driving scenarios due to complex backgrounds, dynamicobjects, and sparse views. We propose AutoSplat, a framework employing Gaussiansplatting to achieve highly realistic reconstructions of autonomous drivingscenes. By imposing geometric constraints on Gaussians representing the roadand sky regions, our method enables multi-view consistent simulation ofchallenging scenarios including lane changes. Leveraging 3D templates, weintroduce a reflected Gaussian consistency constraint to supervise both thevisible and unseen side of foreground objects. Moreover, to model the dynamicappearance of foreground objects, we estimate residual spherical harmonics foreach foreground Gaussian. Extensive experiments on Pandaset and KITTIdemonstrate that AutoSplat outperforms state-of-the-art methods in scenereconstruction and novel view synthesis across diverse driving scenarios. Visitour project page at https://autosplat.github.io/.