Voxel Graph Attention for 3-D Object Detection From Point Clouds
Bin Lu, Sun Ok Yang, Zhenyu Yang · IEEE Transactions on Instrumentation and Measurement · 2023
Recently, 3D object detection from LiDAR point clouds has advanced rapidly, but the detection performance still has shortcomings in complex scenarios. In this paper, we introduce a novel two-stage 3D object detector that utilizes only raw point clouds as input. Specifically, we propose two novel strategies that can effectively enhance the detection performance: the Feature Interpolation Aggregation (FIA) and Graph Attention Pooling (GAP). The former is used to initialize the graph centroid features across different resolutions, which is essential for constructing local graphs. The latter constructs hierarchical local graphs on multi-scale voxels to extract region of interest (RoI) features by considering the contributions of neighboring voxels. This enables the aggregation of more effective and robust features for detection. The proposed innovations enhance the detector’s feature extraction capability and improve its robustness in complex scenarios. Experimental results on the KITTI and Waymo datasets demonstrate that VoxelGraph-RCNN achieves competitive performance compared with state-of-the-art detectors. Specifically, the proposed model surpasses the baseline Voxel R-CNN by 0.43%, 1.04%, and 0.87% in detecting objects with easy, moderate, and hard difficulty levels on the KITTI datasets, respectively.