Virtual Negative Sample Generation Based on Unreliable Pseudo-Labels for Semi-Supervised Semantic Segmentation

Semi-Supervised semantic segmentation primarily involves pseudo-labeling and consistency regularization. Pseudo-Labeling creates labels for unlabeled data to provide additional supervision, while consistency regularization ensures model predictions remain consistent across different views to enhance robustness. Both methods depend on the quality of pseudo-labels, which can significantly impact performance. To avoid discarding data due to low-confidence pseudo-labels, this work identifies that such unreliable pseudo-labels often exhibit competition among top classes but still retain discriminative information for lower-probability classes. Based on this observation, contrastive learning is introduced by treating unreliable pixels as negative samples for specific categories, thereby enabling more effective utilization of unlabeled data. Furthermore, standard contrastive learning typically relies on randomly sampled negative examples, which may include unrepresentative samples or outliers. To address this issue, Virtual Negative Sample Generation (VNSG) is proposed, which leverages global information to generate virtual negative samples positioned near decision boundaries. These samples serve to push anchors and negatives further apart in the feature space, thereby enhancing the effectiveness of contrastive learning. Experimental results demonstrate that the proposed approach achieves competitive performance on the PASCAL VOC 2012 and Cityscapes datasets.