Abstract
Reconstructing dynamic 3D urban scenes for autonomous driving requires balancing fidelity and efficiency. Existing methods are typically semantically agnostic and allocate resources uniformly across scene elements, wasting computation on non-critical regions. We propose Priority-Adaptive Gaussian Splatting (PAGS), which injects task-aware semantic priorities into both reconstruction and rendering. PAGS introduces semantically guided pruning and regularization to preserve detail on safety-critical objects while simplifying less important regions, and a priority-driven rendering pipeline with depth pre-pass culling. Experiments on Waymo and KITTI demonstrate strong reconstruction quality on critical objects, significantly faster training, and rendering speeds above 350 FPS.