CMFST : Class-based Multi-scale Fusion Self-training for Adapting Semantic Segmentation
Gang Zhang, Yinan Ma, Jing Bo Wu, Chengnian Long · 2022
Learning domain-invariant representations has been adopted by mainstream approaches of unsupervised domain adaptation (UDA) to alleviate the domain gap between source data with rich label and unlabeled target data. Self-training is a simple and popular method in UDA by training networks with target pseudo-labels. However, the pseudo-labels’ quality is limited by the single scale output of unadapted network, which has an important impact on training process. To address this issue, we propose a class-based multi-scale fusion (CMF) method that allows the unadapted network to achieve better prediction results by confirming class-scale relation and fusing specific classes prediction from appropriate scales. Our approach doesn’t rely on complex components, therefore is simple, efficient, and easy to transfer. We achieve competitive performance (measured by the Intersection-over-Union, IoU) in the GTA5 to Cityscapes and the SYNTHIA to Cityscapes scenarios which demonstrates the effectiveness of our method.