Using Efficient IoU loss function in PointPillars Network For Detecting 3D Object

Detecting three-dimensional (3D) objects has attracted the growing attention in 3D computer vision research. However, the low precision value is a significant trouble in many applications, like automatic driving, robotics, and medical applications. To solve the low precision problem, we use the 3D EIoU loss as localization loss, which emphasizes on the overlapping degree, central position, and structural shape between two rectangular bounding boxes. Furthermore, we propose an EIoU-NMS to enhance the process of suppressing redundant detecting boxes. By incorporating the 3D EIoU loss and EIoU-NMS into the PointPillars one-stage detectors, the detection performance for 3D point cloud objects is considerably improved. By using the KITTI benchmark, empirical experiments have been conducted to measure the average precision (AP) values for detecting Car, Cyclist, and Pedestrian objects.