Blind Domain Adaptation For Semantic Segmentation via Edge Attention

Ali Solak, Cihan Topal

Abstract: Domain adaptation is a special type of transfer learning that aims to train machine learning architectures trained on a dataset to work on data created for the same task but with a different distribution. Blind domain adaptation is when only the source domain data is accessible during training and the target domain is unknown. In this study, we propose an edge attention module for the semantic segmentation problem to enable the model trained on synthetic datasets to work on real images in the target domain. Since there is no access to the target domain's data distribution in the blind domain adaptation, it is aimed to let the network focus the edges that will be common to both domains through the attention mechanism. Experiments show that the proposed method improves the segmentation performance of the fully convolutional network up to %27.1.

External IDs:dblp:conf/siu/SolakT23