Decoupling Classification and Localization for Single Stage 3D Vehicle Detection from Point Clouds

Abstract

The single shared head detector is mostly used for both classification and regression tasks in object detection networks. However, the two tasks focus on different features, and the features extracted from a single head may inhibit each other’s expression. In 2D detection, a decoupled representation with multi-head structure works better than the shared one for both tasks in the R-CNN [1] blocks. Since the RCNN structure is not available for single-stage detectors, the decoupling must be implemented in a slightly different way. Therefore, we propose a double-head structure based on single-stage point cloud object detection, with two decoupled heads to describe the classification and regression tasks in the high-level feature map, respectively. One of the heads is used to extract classification information, while the other one is used for regression information. In particular, the regression information also contributes to the classification task, and we denote the category result by the union of the classification values predicted by both the regression head and the classification head. Our approach can improve the AP of the baseline in the car detection task on the kitti dataset.