PaFi-GS: Gaussian Splatting via Propagation-Aware Filtering for Urban Street View Rendering

Ying Long, Zhiliu Yang, Hongyu Chen, Zhiyong Hao, Chen Liu

Abstract: Recently, scene representation based on the 3D Gaussian Splatting (3DGS) significantly boost the quality and speed for urban scene rendering compared to methods based on neural radiance fields. However, most of the existing 3DGS-based methods suffer from rendering quality degradation caused by the sparse initialization and weak textured surfaces. Hence, this paper proposes a novel 3DGS-based rendering framework named PaFi-GS. Firstly, an incremental propagation strategy is used to optimize the point cloud on the weakly textured surface, waiving the system from extra point cloud initialization. Next, geometric adaptive filtering and prior reconstruction knowledge are used to create Gaussian points. Then, a hybrid anti-aliasing technique is introduced to adapt to different levels of geometry details, significantly improving the anti-aliasing effect of unbounded scenes. Finally, a CNN-based appearance decoupling module and the multi-loss function are exploited to foster the model to handle illumination variance. Extensive experiment results on Waymo and KITTI datasets demonstrate the effectiveness and scalability of the proposed method.

External IDs:dblp:conf/icann/LongYCHL25