Point-HN: Unified spatial-context modeling for fast and accurate 3D point cloud detection

Point cloud object detection is a fundamental task in various 3D perception applications such as autonomous driving and mobile robotics. However, most existing methods suffer from inaccurate 3D localization and intensive computational costs due to the inherent sparsity and irregular spatial distribution of point cloud data. This paper presents Point-HN, a point–voxel Hybrid Network that effectively balances accuracy and efficiency in 3D object detection on point clouds. Point-HN incorporates two designs that jointly model spatial structures and preserve fine-grained geometric details. The first is a Retentive Spatial Transformer (RST) that explicitly models spatial relations between points and voxels, thereby highlighting critical features around potential object regions. The second is a Geometric-Context Enhancer (GCE) that introduces a graph-based representation to capture geometric relations among key points, thereby effectively compensating for the loss of fine-grained information during voxelization with minimal overhead. Extensive experiments across multiple widely adopted autonomous driving benchmarks show that Point-HN achieves superior performance in 3D object detection consistently, underscoring its broad applicability in various real-world 3D perception tasks.