Unsupervised Domain Adaptive Semantic Segmentation with Alternate Guidance from Multiple Teachers

In unsupervised domain adaptive semantic segmentation, the Mean Teacher framework has demonstrated strong performance. By utilizing two models with the same structure but different initializations, it effectively leverages unlabeled data to train the model. However, in this framework, the Teacher Model's (T-Model) parameters are updated as the exponential moving average (EMA) of the Student Model's (S-Model). This update mechanism heavily relies on the S-Model and may reinforce erroneous knowledge if the S-Model learns incorrect information. We propose an Alternating Guidance from Multiple Teachers (AGMT) method to address the issues inherent in the Mean Teacher framework. Our approach involves initializing three T-Models with the same network but different parameters at the start of training. During training, these T-Models are dynamically activated in an alternating manner to guide the training of the S-Model. Moreover, we introduce correlation-mixed sampling and boundary sampling techniques, which help the model learn better boundary features, thereby improving segmentation accuracy for rare classes.