Avoid Orientation Confusion in Symmetrical Oriented Object Detection

With development of the deep learning methods for 3D object detection, various solutions have been proposed based on the LiDAR point clouds. Since the point cloud only present object contours, it’s difficult for network to distinguish completely opposite orientations of the object when facing the point cloud with highly symmetrical shape, which could lead to jitter for the predicted bounding box. In this paper, we propose multiple strategies, including pillar-level supervision and multi-phase orientation encoder inspired by Phase-Shifting Coder [1], as auxiliary modules. They are only utilized in training process. The encoded orientation value with multiple phases unifies two opposite directions of the object’s symmetry axis and distinguish the directions along the asymmetric axis at the same time. Thus, it solves the fuzzy predicting problem on the defined rotation boundaries when the network faces the same input. By mapping the ground truth to the feature map in pillar-level supervision module, the foreground pixels within the instance are prominently learned by the network to stable the prediction of the objects’ location and orientation. We adopt these strategies simultaneously in closed operating port to achieve an expected performance.