Go to ICLR 2024 Conference homepageCT++: Complementary Co-Training for Semi-Supervised Semantic Segmentation
With limited annotations, semi-supervised semantic segmentation aims to enhance the segmentation ability through abundant unlabeled images. Among recent trends, co-training is gaining increasing popularity, where two parallel models produce pseudo labels for each other. The success of co-training heavily relies on the discrepancy of peer models. To achieve this, prior works mostly leverage different initializations in decoders. Unfortunately, the two models still quickly converge to an extremely coupling state, making co-training downgrade to poorer self-training. To address this dilemma, we present our CT++, decoupling dual Co-Training models from two novel perspectives. First, we propose to construct complementary feature-level views. We design two co-training models to utilize disjoint and complementary sets of features for decoding. Apart from complementary features, we further seek complementary input views for the two models to learn respectively. Our two complementary principles enlarge the model discrepancy significantly, enabling co-training models to transfer distinct knowledge to each other and broaden their capability. This contributes to remarkably boosted co-training effectiveness. Extensive studies on Pascal, Cityscapes, COCO, and ADE20K exhibit the strong superiority of our method, e.g., 80.2% mIoU with only 92 labels on Pascal.