Improving pseudo labels with intra-class similarity for unsupervised domain adaptation

Unsupervised domain adaptation (UDA) transfers knowledge from a label-rich source to different but related fully-unlabeled target domain. To address the problem of shift, more and UDA methods adopt pseudo labels samples improve generalization ability on However, inaccurate may yield suboptimal performance with error accumulation during optimization process. Moreover, once are generated, how remedy generated is far explored. In this paper, we propose novel approach accuracy in It first generates coarse by conventional method. Then, it iteratively exploits intra-class similarity for improving labels, aligns domains improved labels. The improvement made deleting dissimilar samples, then using spanning trees eliminate wrong samples. We have applied proposed several as an additional term. Experimental results demonstrate that method can boost further lead discriminative invariant features than baselines.