DrivingVoxels: Compositional Sparse Voxel Rasterization for Dynamic Driving Scene Reconstruction

2026-06-22

AI Summary

This work addresses the challenges of dynamic driving scene reconstruction, which include unbounded environments, coexisting multiple dynamic objects, and the difficulty of existing methods to balance efficiency and accuracy. The authors propose a compositional sparse voxel rendering framework that introduces multi-octree structures for the first time in this task: the static background and each rigidly moving object are modeled independently using sparse octrees within their respective local coordinate systems, enabling efficient geometry reconstruction through LiDAR-guided initialization. All components are jointly rendered in a single rasterization pass, yielding an explicit representation without requiring neural networks. Experiments on PandaSet demonstrate that the method significantly improves geometric fidelity while maintaining perceptual quality, with substantially lower training time and memory consumption compared to state-of-the-art approaches such as 3D Gaussian Splatting.

Abstract

Reconstructing dynamic urban scenes remains challenging due to the unbounded nature of driving environments and the presence of multiple dynamic objects. Currently, potentially faster sparse voxel methods are mainly designed for static scenarios. On the other hand, dynamic approaches based on 3D Gaussian Splatting, despite their high-fidelity, are often time-consuming for driving scenarios and exhibit uncontrollable memory growth in large scenes. To address these limitations, we present DrivingVoxels, a compositional sparse voxel rendering framework for dynamic driving scenes. Our method jointly rasterizes sparse voxels from multiple independent octrees within a single rendering pass. Each rigid dynamic object is represented by an octree defined in its local coordinate frame, while a separate static octree models the stationary background. DrivingVoxels adopts a fully explicit, neural-free representation together with a LiDAR-guided structural initialization that efficiently captures scene geometry. We evaluate our framework on the PandaSet benchmark, demonstrating that DrivingVoxels performs on par on perceptual metrics and better on structural metrics for NVS and reconstruction while requiring shorter training times than previous 3DGS-base methods to an efficient optimization workflow anchored by a strong LiDAR prior.

Problem

• dynamic driving scene reconstruction
• sparse voxel
• 3D Gaussian Splatting
• memory efficiency
• unbounded urban environments

Innovation

• Compositional Sparse Voxels
• Dynamic Scene Reconstruction
• Octree-based Rendering
• LiDAR-guided Initialization
• Neural-free Representation