SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction

Vision-based 3D semantic occupancy prediction is critical for autonomous driving, yet existing dense representations are computationally expensive, while projections like BEV or TPV lose fine-grained structural information. The authors investigate whether a fully sparse representation can achieve a better balance between structural fidelity and computational efficiency by explicitly modeling the geometry of the scene.

SparseOcc++ introduces a geometry-aware sparse representation that separates scene completion from semantic classification. Instead of indiscriminately propagating features into empty space, the model uses a coarse-to-fine strategy to generate sparse anchor voxels. It then learns a Scene Completion Field (SCF) that regresses signed distances to scene boundaries. To handle the anisotropy of outdoor scenes, the authors employ an orthogonal decomposition strategy, splitting the distance prediction into planar (X-Y) and vertical (Z) components. A geometry-guided propagation mechanism then uses these predicted distances to complete the scene, ensuring that subsequent semantic segmentation is performed only on geometrically verified occupied regions.

SparseOcc++ establishes a new state-of-the-art in occupancy prediction. On the nuScenes dataset, it improves IoU by 2.3 points and runs 3.9x faster than the previous SparseOcc method. On SemanticKITTI, it achieves a 5.9x speedup compared to the dense OccFormer baseline while maintaining superior accuracy. The geometry-guided propagation effectively eliminates trailing noise and hallucinations common in dense methods, particularly for small objects like pedestrians and cyclists.

This work demonstrates that 3D perception can be both highly accurate and computationally efficient by leveraging the inherent sparsity of 3D environments. By reformulating scene completion as a geometric regression task rather than a dense classification task, the framework avoids the computational waste of processing empty space, making it a highly scalable solution for real-time autonomous driving systems.