An Efficient Intersection Over Union Loss Function for 3D Object Detection

In the area of computer vision, object detection using convolutional neural networks (CNNs) has become quite a popular procedure because of their effectiveness and simplicity. The loss function has a great influence on the average accuracy value of the CNN model’s detector results. An improved 3D efficient intersection over union (EIoU) loss function is proposed to improve the localization accuracy. The diagonal distance between bounding boxes’ corners and centers, with the dimensional change in boxes’ geometry sides, are used for matching between the 3D predicted bounding box with the 3D ground truth bounding box. By taking the geometry sides change and diagonal distance between the 3D predicted and 3D ground truth boxes, a great influence on the localization accuracy is generated. For the network model, the localization accuracy is improved because of the strength of diagonal distance and geometry sides’ adjustment. Utilizing the one-stage object detector 3D YOLO v4 and applying the 3D EIoU experimentally on the KITTI dataset, the findings demonstrate the effectiveness of the 3D EIoU in improving the accuracy of localization for the network model. Compared with 3D GIoU, the proposed EIoU enhances the average precision (AP) value by 0.24% and AP70 by 1.179% in the car class, AP by 0.578%, and AP55 by 6.022% in cyclist class, and AP by 0.1531% and AP30 by 2.548% in pedestrian class.