SurfaceGS: Dynamic Surface Gaussian Splatting for Urban Driving Scenes

Fudong Ge, Dingning Liu, Hanshi Wang, Yiwei Zhang, Jin Gao, Weiming Hu, Zhipeng Zhang

Abstract: Reconstructing dynamic scenes with 3D Gaussian Splatting is fundamentally challenged by motion and geometric inconsistencies, which often lead to visual artifacts like unnatural stretching or temporal flickering. These issues stem from the unconstrained nature of individual Gaussians, which lack an explicit underlying surface to guide their movement. To address this, we introduce SurfaceGS, a novel paradigm that structurally embeds geometric priors into the reconstruction process through a dual-layer surface-Gaussian representation. This core innovation explicitly anchors the radiance-carrying Gaussians to a set of learnable, dynamic surfaces, ensuring that they move and deform coherently with the object's true geometry. We model complex scene dynamics via a hybrid approach, using Bézier curves for global trajectories and efficient B-spline surfaces for local, non-rigid deformations. To ensure scalability for large-scale scenes, we introduce a suite of optimizations, including a patch-wise decomposition and a new adaptive control mechanism tailored for our dynamic surface paradigm. Extensive experiments on challenging urban driving benchmarks validate our approach, demonstrating that SurfaceGS achieves new state-of-the-art results in photorealistic rendering. In addition, our method enables controllable simulation, such as editing and extrapolation for autonomous driving scenes, facilitating safe self-driving systems research.