Twin Pseudo-training for semi-supervised semantic segmentation

Recent advancements in semi-supervised semantic segmentation have demonstrated the effectiveness of utilizing pseudo labels supervision to mitigate the limitations of pixel-wise annotations. However, pseudo-labels generated using self-training techniques typically contain a significant amount of noise, which can impede the training process of the supervised model. In this study, we identify low- and high-level semantic errors as the two key factors that hinder the accuracy of pseudo labels. To fully exploit the potential of pseudo labels, we introduce a novel semi-supervised framework named Twin Pseudo-training (TPseudo), which employs a consistency and disagreement collaboration strategy. Specifically, we correct pseudo labels with a False-positive Filter (FPF) to reduce high-level semantic noise and refine low-level semantic biases using a Semantic Error Detector (SED). Lastly, we design a Self-Adaptive Weight (SAW) loss function based on a disagreement between two predictions to exploit each pixel of pseudo labels. Experimental results on the standard benchmarks PASCAL VOC2012 and Cityscapes demonstrate the efficacy of the proposed method.