Roy Subhankar ; Krivosheev Evgeny ; Zhong Zhun ; Sebe Nicu ; Ricci Elisa

Abstract

In this paper we address multi-target domain adaptation (MTDA), where givenone labeled source dataset and multiple unlabeled target datasets that differin data distributions, the task is to learn a robust predictor for all thetarget domains. We identify two key aspects that can help to alleviate multipledomain-shifts in the MTDA: feature aggregation and curriculum learning. To thisend, we propose Curriculum Graph Co-Teaching (CGCT) that uses a dual classifierhead, with one of them being a graph convolutional network (GCN) whichaggregates features from similar samples across the domains. To prevent theclassifiers from over-fitting on its own noisy pseudo-labels we develop aco-teaching strategy with the dual classifier head that is assisted bycurriculum learning to obtain more reliable pseudo-labels. Furthermore, whenthe domain labels are available, we propose Domain-aware Curriculum Learning(DCL), a sequential adaptation strategy that first adapts on the easier targetdomains, followed by the harder ones. We experimentally demonstrate theeffectiveness of our proposed frameworks on several benchmarks and advance thestate-of-the-art in the MTDA by large margins (e.g. +5.6% on the DomainNet).