PSA-Det3D: Pillar set abstraction for 3D object detection

Small object detection for 3D point cloud is a challenging problem because of two limitations: (1) The sparsity clouds significantly increases the difficulty perceiving small objects. (2) occlusion objects can easily break shape their clouds. To alleviate these problems, we design point-based network PSA-Det3D which mainly consists pillar set abstraction (PSA) and foreground compensation (FPC). PSA improves query approach abstraction, benefits point-wise feature aggregation FPC fuses points estimated centers to select candidate points, effectively performance occluded Extensive experiments show that our proposed achieves higher on all categories. For detection, method outperforms existing based algorithms KITTI benchmark.