Bilateral-Branch Fusion Network for 3D Object Detection

In this article, we focus on exploring the 3D object detection of LIDAR-RGB fusion, but there are still two challenging problems that need to be solved: 1) The lack of information of distant point cloud samples leads to misalignment of multi-sensor data fusion. 2) When the center points coincide and the aspect ratio remains the same, the loss based on the traditional iou and giou remains unchanged, resulting in a deviation in the back propagation. In this work, we propose a multi-mode data Bilateral-Branch Network (MBN), which includes a Bilateral-Branch Network (BBN) and a new 3d iou loss to solve these two problems. Specifically, we designed two branches of representation learning and classifier learning, called the regular learning branch and the rebalancing branch, respectively. The two branches use the same network structure. These two branches use parameters adjusted according to the number of trained epochs to adaptively adjust the entire model. In addition, our 3D IoU loss calculates the ratio of the length, width and height of the anchor box to avoid simplistic optimization of non-overlapping bounding boxes. A large number of experiments in the KITTI benchmark test prove that our performance is superior to the most advanced methods.