Domain generalization aims to help models trained on a set of sourcedomains generalize better on unseen target domains . We propose a Domain-Irrelevant UnsupervisedLearning (DIUL) method to cope with the significant and misleadingheterogeneity within unlabeled data . DIUL can not onlycounterbalance the scarcity of labeled data but also strengthen thegeneralization ability of models when the labeled data are sufficient . As apretraining approach, DIUL shows superior to ImageNet pretraining protocol evenwhen the available data are unlabeling and of a greatly smaller amount comparedto ImageNet. Extensive experiments clearly demonstrate the effectiveness of ourmethod compared with state-of-the-art unsupervised learning counterparts. We observe that DIUL is superior to the ImageNet training protocol even when the availableData are

Author(s) : Xingxuan Zhang, Linjun Zhou, Renzhe Xu, Peng Cui, Zheyan Shen, Haoxin Liu

Links : PDF - Abstract

Code :

Keywords : data - domain - diul - imagenet - models -

Leave a Reply

Your email address will not be published. Required fields are marked *