How to enrich cross-domain representations? Data augmentation, cycle-pseudo labeling, and category-aware graph learning

Ba Hung Ngo, Doanh C. Bui, Tae Jong Choi

Abstract: Highlights•We first identify factors impacting semantic representations of input objects.•Data augmentation, cycle pseudo-labeling, and graph learning enrich features.•We unify these factors into an end-to-end framework called EnSR.•EnSR achieves state-of-the-art results without added inference complexity.