Boundary-Aware Set Abstraction for 3D Object Detection

The basic components of a point-based 3D object detector is set abstraction (SA) layer, which down-samples points for better efficiency and enlarges receptive fields. However, existing SA layer only takes the relative locations among points into consideration, e.g. using furthest point sampling (FPS), while ignoring point features. Because the points on the objects take small proportion of space, the cascaded SA may miss to contain objects' points in the last layer, which will degrade the 3D object detection performances. We design a new lightweight and effective SA layer named Boundary-Aware Set Abstraction layer (BA-Net) to retain important foreground and boundary points during cascaded down-sampling. Technically, a lightweight point segmentation model (PSM) to compute the point-wise foreground scores is firstly embedded, then a Boundary Prediction Model (BPM) to detect the points on the object boundaries is proposed. These semantic scores are used to weight inter-node distances and the Boundary-aware Furthest Point down-Sampling (B-FPS) is conducted in this twisted distance space. Experimental results show that BA-Net can enhance the average accuracy (mAP) of the most competitive car class on the KITTI dataset by 1.14% and 1.06% in two different baseline. Code is available at https://github.com/HuangZhe885/Boundary-Aware-SA.