Graph-Guided Feature Fusion and Group Contrastive Learning for Domain Adap-tation Semantic Segmentation
Considering the problem of unsupervised domain adaption semantic segmentation,it is very important to establish a long-distance context relationship between the source domain and the target domain and how to solve the problem of unbalance distribution of different classes of pixels.we propose a dual cross-domain graph convolution network to exploit the long-distance context between source and target domain and fuse the feature of two domains.Specifically,we construct the position similarity matrix and channel similarity matrix of the cross domain and propose the cross-domain position graph convolution and cross-domain channel graph convolution.In order to solve the problem of unbalanced distribution of classes in the datasets and capture more domain invariant feature,we propose a group contrastive learning strategy to narrow the distance between the same class of two domains and widen the distance between the different classes of two domains by constructing positive and negative samples in the group.A large number of experiments show that our method achieves good performance on Urban Scene datasets GTA5 to Cityscapes and SYNTHA to Cityscapes.
万方数据
维普
