3-D Object Detection With Balanced Prediction Based on Contrastive Point Loss

Object detection on point clouds is widely used in autonomous driving technology. Recent researches have demonstrated that good feature representation is the key to 3D object detection, especially for point-based methods. Meanwhile, contrastive learning has shown its effectiveness to learn the general visual representation of images as a pre-training paradigm. Motivated by this, the contrastive learning is extended to the task of object detection for better distinguishing different types of objects. In this paper, a simple and effective single-stage detector, named 3D Object Detection with Balanced Prediction based on Contrastive Point Loss (BP-CPL), is proposed with a Contrastive Point Loss and a Re-balanced branch. Through the Contrastive Point Loss, similar representations among points of the same category and discriminative information among points of different categories can be learned. In order to keep the contrastive points consistency, a point filter is proposed. In addition, compared to the original point-wise feature, the self-supervised learned embedding is more robust to few-sample categories. Therefore, a re-balanced branch combined with the origin classification branch in a cumulative learning manner is proposed to re-balance the prediction results during the training phase. Extensive experiments show the effectiveness of the proposed method, especially for few-sample objects. The code will be available at https://github.com/Tongjiaxun/BP-CPL.