Semi-supervised Semantic Segmentation with Uncertainty-Guided Self Cross Supervision

As a powerful way of realizing semi-supervised segmentation, the cross supervision method learns cross consistency based on independent ensemble models using abundant unlabeled images. In this work, we propose a novel cross supervision method, namely uncertainty-guided self cross supervision (USCS). To avoid multiplying the cost of computation resources caused by ensemble models, we first design a multi-input multi-output (MIMO) segmentation model which can generate multiple outputs with the shared model. The self cross supervision is imposed over the results from one MIMO model, heavily saving the cost of parameters and calculations. On the other hand, to further alleviate the large noise in pseudo labels caused by insufficient representation ability of the MIMO model, we employ uncertainty as guided information to encourage the model to focus on the high confident regions of pseudo labels and mitigate the effects of wrong pseudo labeling in self cross supervision, improving the performance of the segmentation model. Extensive experiments show that our method achieves state-of-the-art performance while saving 40.5% and 49.1% cost on parameters and calculations.