Multi-Target Domain Adaptation with Class-Wise Attribute Transfer in Semantic Segmentation

Abstract

In this paper, we present a novel multi-target domain adaptation (MTDA) method that adapts a single model to multiple domains with class-wise attribute transfer. To achieve this, we propose a high-precision pseudo labeling method for target domain images by utilizing cross-domain correspondence matching, which matches a target region to the most similar source region. Then, we propose class-wise image translation using the pseudo labels to avoid the problem of transferring characteristics between different classes and to allow translation between the same classes. Lastly, we introduce cross-domain feature consistency to learn the different characteristics of each target domain. Extensive experiments on the various complex driving scene show that ours achieves better performance than other state-of-the-art methods. The dense ablation study demonstrates the effectiveness of the proposed method. © 2023. The copyright of this document resides with its authors. It may be distributed unchanged freely in print or electronic forms.