Information Transfer in Semi-Supervised Semantic Segmentation

Enhancing the accuracy of dense classification with limited labeled data and abundant unlabeled data, known as semi-supervised semantic segmentation, is an essential task in vision comprehension. Due to the lack of annotation in unlabeled data, additional pseudo-supervised signals, typically pseudo-labeling, are required to improve the performance. Although effective, these methods fail to consider the internal representation of neural networks and the inherent class-imbalance in dense samples. In this work, we propose an information transfer theory, which establishes a theoretical relationship between shallow and deep representations. We further apply this theory at both the semantic and pixel levels, referred to as IIT-SP, to align different types of information. The proposed IIT-SP optimizes shallow representations to match the target representation required for segmentation. This limits the upper bound of deep representations to enhance segmentation performance. We also propose a momentum-based Cluster-State bar that updates class status online, along with a HardClassMix augmentation and a loss weighting technique to address class imbalance issues based on it. The effectiveness of the proposed method is demonstrated through comparative experiments on PASCAL VOC and Cityscapes benchmarks, where the proposed IIT-SP achieves state-of-the-art performance, reaching mIoU of 68.34% with only 2% labeled data on PASCAL VOC and mIoU of 64.20% with only 12.5% labeled data on Cityscapes.