Title: SparseStreet: Enabling Real-Time Street Scene Simulation via Sparse Gaussian Splatting

Abstract:

Although 3D Gaussian Splatting has demonstrated significant potential in reconstructing street environments, current approaches typically rely on an excessive volume of Gaussian primitives to capture intricate details. This reliance results in rendering speeds that are too sluggish and storage requirements that are prohibitively high. We note that dynamic elements, such as pedestrians and vehicles, necessitate high-fidelity modeling to ensure temporal coherence, whereas static background areas are often characterized by considerable redundancy. To address this disparity, we present SparseStreet, a specialized compression framework tailored for street scenes. Our approach first employs a node-based, learnable pruning mechanism that methodically eliminates low-impact Gaussian primitives while safeguarding visually essential regions. Once the scene representation reaches stability, we implement a background compression stage to further minimize redundancy within static zones. This strategy successfully maintains the geometric and visual integrity of moving objects while drastically lowering the count of Gaussian primitives. Comprehensive evaluations on the Waymo and nuScenes datasets indicate that SparseStreet delivers a compression ratio of up to 80% with negligible impact on quality, thereby facilitating resource-efficient, high-fidelity reconstruction of dynamic scenes.

Project website: https://sparsestreet.github.io/.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC