Weakly Supervised Semantic Segmentation for Point Clouds with Self-Purification on Pseudo Labels

Currently, research in the field about point cloud semantic segmentation primarily focuses on fully supervised learning, which requires expensive manual point-level annotations. Weakly supervised learning is an approach to overcome the time and effort required for such annotations. However, for large-scale point clouds with limited labeled points, it is challenging for networks to learn the differences between point features. Some methods address this issue by pretraining the network with generated pseudo-labels for the point cloud, allowing the network to learn more information. However, pseudo-labels are prone to a certain proportion of mislabeling, and directly using them can lead to erroneous classification results. To address this problem, we propose a pseudo-label self-purification framework. Specifically, we design confidence scores to dynamically select reliable labeled points and employ clustering techniques to partially update the pseudo-labels, thereby providing correct supervision for network learning. Additionally, inspired by unsupervised learning, we introduce a perturbation branch to enhance the consistency of point cloud predictions. The experimental results, obtained from the evaluation of two extensive datasets, demonstrate a gain of 2.0% on average compared to recent weakly supervised method. Furthermore, our approach achieves results that are comparable to certain fully supervised methods.