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

Abstract

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