A semi-supervised 3D object detection method for autonomous driving

Abstract

3D object detection is very important for autonomous driving, but most of algorithms heavily rely on data annotation, which is a laborious job. Semi-supervised learning is an effective strategy to reduce the heavy annotation burden in this task. In this paper, we propose a semi-supervised 3D object detection method. To reduce the workload of 3D annotations, we adopt the teacher-student framework to generate pseudo-labels from unlabeled training data, and use a label filtering method to improve the pseudo label quality. For the convenience of deployment, we introduce the simple and fast one-stage object detector. We validate our method on the KITTI dataset. For moderate Car detection task, our method can achieve 76.28 mAP using half labels compared with 77.34 mAP of the PointPillars using all labels. Experiment results show that our method can also achieve approximate performance on Pedestrian and Cyclist detection task with fewer labels.