PSA-Det3D: Pillar Set Abstraction for 3D object Detection

Existing detectors have achieved high accuracy for car category, but the detection performance for small objects such as pedestrian and cyclist is still unsatisfactory. Thus, how to improve the detection of small objects in 3D point cloud is an important yet challenging problem [7]. Some methods apply the multimodal fusionbased approaches in the 3D object detection network [15, 27, 30]. Due to the detailed information provided by the cameras and LiDAR sensors, they achieve high accuracy for both normal and small objects. However, the complex fusion networks increase the computational cost of these methods and thus limit their application. Although the researches on point-based methods [3, 23, 32, 33] have achieved remarkable progress, the limitations of small objects are not sufficiently considered yet. There are two obvious limitations for LiDAR-based small object detection: (1) Perceiving small objects is much more difficult than normal objects because the sparse LiDAR-based point clouds usually do not provide sufficient information of small objects.