A Pseudo Variance Algorithm for Semi-Supervised Semantic Segmentation

The key aspect of semi-supervised semantic segmentation lies in the construction of pseudo-labels. The conventional approach involves selecting the prediction with the highest confidence as the pseudo-label. However, this method can introduce an issue where the class with the highest confidence may be an incorrect prediction, especially for categories with ambiguous features, while another high-confidence prediction may be correct. To address this concern, we propose an effective method for constructing pseudo-labels that integrates two high-confidence sources of information to mitigate the impact of erroneous predictions. Specifically, we devise a pseudo-variance function that combines the two highest confidences to better evaluate the quality of prediction results. Additionally, an adaptive coefficient is set to enable the method to adapt to various datasets. The pseudo variance algorithm is integrated into the training process by incorporating a novel loss function, which maps the prediction results to pseudo-labels associated with their confidences and subsequently re-adjusts the network parameters. Experimental results on different datasets demonstrate that our method outperforms previous approaches under similar conditions, showcasing superior performance.