PSCO: A Point Cloud Scene Classification Model Based on Contrast Learning

Point cloud LiDAR data are increasingly used for detecting road situations for autonomous driving. As data acquisition is much cheaper than data annotation, the classification of the collected data and selecting a part of it for annotation is an important pre-processing step. In this study, we propose a self-supervised method and a corresponding package (PSCO) to classify driving scenes in large-scale point cloud data. We propose a contrastive learning approach which encodes each single frame of the collected raw data into a feature vector that can be effectively classified. Using this approach, the number of necessary frames for annotation to train scene-based models can be significantly reduced. In addition, our method can also select specific scenario data for different individual training stages.