Boosting Semi-Supervised Semantic Segmentation with Probabilistic Representations

Recent breakthroughs in semi-supervised semantic segmentation have been developed through contrastive learning. In prevalent pixel-wise learning solutions, the model maps pixels to deterministic representations and regularizes them latent space. However, there exist inaccurate pseudo-labels which map ambiguous of wrong classes due limited cognitive ability model. this paper, we define from a new perspective probability theory propose Probabilistic Representation Contrastive Learning (PRCL) framework that improves representation quality by taking its into consideration. Through modelling mapping as via multivariate Gaussian distributions, can tune contribution tolerate risk pseudo-labels. Furthermore, prototypes form indicates confidence class, while point prototype cannot. More- over, regularize distribution variance enhance reliability representations. Taking advantage these benefits, high-quality feature be derived space, thereby performance se- mantic further improved. We conduct sufficient experiment evaluate PRCL on Pascal VOC CityScapes demonstrate superiority. The code is available at https://github.com/Haoyu-Xie/PRCL.