Reliable Domain Adaptation With Classifiers Competition for Image Classification

摘要

Unsupervised domain adaptation (UDA) tries to utilize the labeled source domain knowledge to help the learning of unlabeled target domain. Existing methods address this problem either by minimizing joint distribution divergence and generating the pseudo target labels by source classifier, or by aligning two domains in manifold subspace. However, they ignore two significant issues: 1) unreliable distribution alignment, which means the source classifier always misclassifies partial target data which may deteriorate adaptation performance when using pseudo labels, 2) insufficient divergence reduction, which means that distribution alignment often focuses on reducing domain shift in original space, where large discrepancy and feature distortion are hard to overcome. On the other hand, reducing distribution divergence only in manifold space is often not sufficient. To alleviate these issues, a Reliable Domain Adaptation (RDA) method is proposed in this brief. Specifically, double task-classifiers and dual domain-specific projections are introduced to align easily misclassified and unreliable target samples into reliable ones in an adversarial manner. Moreover, the domain differences in both manifold and category space are eliminated. Extensive experiments on diverse databases prove the effectiveness of RDA over state-of-the-art unsupervised domain adaptation methods.