Consistency-Based Semi-Supervised Learning for Point Cloud Classification

Under the strong supervision information of a large number of training samples with ground-truth labels, the point cloud classification techniques have achieved great success. However, in real tasks, it is difficult to get strong supervision information like fully ground-truth labels due to the high cost of the data labeling process. Thus, it is desirable to work with weak supervision. This paper presents a semi-supervised learning framework based on consistency constraints on the point cloud to solve the classification task. The key to the framework mainly includes consistency constraints and pseudo-label generation. The former uses consistency constraints to encourage the model to output consistent predictions on the original and the perturbed point cloud, which avoids overfitting due to limited labeled data. The latter generates high-quality pseudo-labels for unlabeled point clouds to strengthen the discriminative supervision of the model. By fully mining the labeled and unlabeled point clouds, the generalization ability of the model is improved. On the 3D point cloud benchmark ModelNet40 [19], with only 10% labeled data, our proposed approach performs comparable results to the supervised approaches.