MADAN: Multi-source Adversarial Domain Aggregation Network for Domain Adaptation

Domain adaptation aims to learn a transferable model bridge the domain shift between one labeled source and another sparsely or unlabeled target domain. Since data may be collected from multiple sources, multi-source (MDA) has attracted increasing attention. Recent MDA methods do not consider pixel-level alignment sources misalignment across different sources. In this paper, we propose novel framework address these challenges. Specifically, design Multi-source Adversarial Aggregation Network (MADAN). First, an adapted is generated for each with dynamic semantic consistency while aligning towards at cycle-consistently. Second, sub-domain aggregation discriminator cross-domain cycle are proposed make domains more closely aggregated. Finally, feature-level performed aggregated training task network. For segmentation adaptation, further enforce category-level incorporate multi-scale image generation, which constitutes MADAN+. We conduct extensive experiments on digit recognition, object classification, simulation-to-real tasks. The results demonstrate that MADAN MADAN+ models outperform state-of-the-art approaches by large margin.