Semantic Occupancy Prediction with Dual Range-Voxel Representation

Autonomous driving systems require accurate 3D semantic occupancy prediction to understand the environment. Existing LiDAR-based methods typically stack multiple sweeps of point clouds to overcome the sparsity and incompleteness of single-sweep data. However, this multi-sweep strategy introduces significant computational overhead and sensitivity to pose transformation noise. The authors investigate whether a single-sweep approach can achieve comparable or superior performance by effectively combining different representations of the scene.

The authors propose a Dual Range-Voxel Representation (DRVR) framework. This approach processes single-sweep point clouds through two parallel paths: a range-view encoder that extracts compact contextual information via spherical projection, and a geometry-aware voxel-view encoder that preserves spatial structure through multi-scale voxelization. These two streams are integrated via a collaborative range-voxel fusion module, which utilizes voxel-to-range and range-to-voxel fusion mechanisms to exchange information between the context-rich range-view and the geometry-rich voxel-view. This allows the model to learn a comprehensive semantic occupancy representation without relying on temporal stacking.

DRVR demonstrates superior performance across three major benchmarks: nuScenes-Occupancy, SemanticKITTI, and SemanticPOSS. On the nuScenes-Occupancy dataset, the single-sweep DRVR achieves a 5.4% improvement in mIoU and a 2.1x acceleration in inference speed compared to multi-sweep baselines. Furthermore, the model exhibits significantly higher robustness to pose transformation noise, as it eliminates the need for multi-sweep alignment. Qualitative results confirm that DRVR effectively recovers fine-grained structures and background regions even under sparse input conditions.