Weakly Supervised Learning for Point Cloud Semantic Segmentation With Dual Teacher

Abstract

Point cloud semantic segmentation has achieved considerable progress in the past decade. To alleviate expensive data annotation efforts, weakly supervised learning methods are preferable, and traditional approaches are typically based on siamese neural networks. To enhance the feature learning capability, in this work, we introduce a dual-teacher-guided contrastive learning framework for weakly supervised point cloud semantic segmentation. A dual-teacher framework can reduce sub-network coupling and facilitate feature learning. In addition, a cross-validation approach can filter out low-quality samples, and a pseudo-label correction module can improve the quality of pseudo-labels. Cleaned unlabeled data are used to construct contrastive loss based on the prototypes of each class, which further boost the segmentation performance. Extensive experimental results conducted on the S3DIS, ScanNet-v2, and SemanticKITTI datasets demonstrate that our proposed DCL outperforms state-of-the-art methods.