Hard class rectification for domain adaptation

Domain adaptation (DA) aims to transfer knowledge from a label-rich and related domain (source domain) label-scare (target domain). Pseudo-labeling has recently been widely explored used in DA. However, this line of research is still confined the inaccuracy pseudo labels. In paper, we explore imbalance issue performance among classes in-depth observe that worse performances all are likely further deteriorate pseudo-labeling, which not only harms overall but also restricts application propose novel framework, called Hard Class Rectification (HCRPL), alleviate problem two aspects. First, simple yet effective scheme, named Adaptive Prediction Calibration (APC), calibrate predictions target samples. Then, consider calibrated ones, especially for those belonging hard classes, vulnerable perturbations. To prevent these samples be misclassified easily, introduce Temporal-Ensembling (TE) Self-Ensembling (SE) obtain consistent predictions. The proposed method evaluated on both unsupervised (UDA) semi-supervised (SSDA). Experimental results several real-world cross-domain benchmarks, including ImageCLEF, Office-31, Office+Caltech, Office-Home, substantiate superiority method.