SP-VoxelDet: Channel-Efficient Fully Sparse 3D Object Detection via Partial Convolution

Abstract

Efficient and reliable 3D perception is fundamental for autonomous robotic systems under strict onboard computational constraints. Although sparse convolution exploits spatial sparsity in LiDAR point clouds, existing methods largely overlook channel redundancy, creating unnecessary overhead on resource-limited platforms. We present SP-VoxelDet, a channel-aware lightweight framework for fully sparse 3D object detection. The core idea is to reduce channel-wise redundancy in sparse voxel representations. Its core operator, PC-SubMConv3d, applies submanifold sparse convolution to only a subset of feature channels while directly propagating the remainder, reducing computation without altering the spatial receptive field. We further introduce a stage-wise channel allocation strategy that applies stronger reduction in early geometric layers and restores full-channel computation in deeper semantic layers. A lightweight late-stage sparse multi-scale fusion module uses normalized learnable scale weights to improve cross-scale aggregation with limited overhead. On KITTI, SP-VoxelDet reduces parameters by 31.1% and improves inference speed by 13.2% over a strong fully sparse baseline, with only a 0.13 mAP decrease. On nuScenes validation, SP-VoxelDet achieves 59.53 mAP and 66.38 NDS, with only 0.47 mAP and 0.72 NDS decreases relative to the baseline, showing that the channel-efficient design remains competitive on a larger-scale benchmark. Improved pedestrian detection suggests particular benefits for small and sparse targets. In an embedded study using a Velodyne VLP-16 LiDAR and NVIDIA Jetson Orin within a ROS 2 pipeline, SP-VoxelDet achieves higher end-to-end throughput under identical conditions. These results indicate that channel-aware sparse computation is effective for resource-constrained LiDAR perception.