Unified Optimal Transport Framework for Universal Domain Adaptation (Supplementary Material) Wanxing Chang Ye Shi 1,3 Jingya Wang

Neural Information Processing Systems 

A.1 Evaluation metric H-score was proposed in CMU [4], which emphasizes the importance of both abilities of UniDA methods. We follow the dataset split in [4, 10] to conduct experiments. We report error bars after running the experiments three times for UniOT. S2, we report the averaged H-score results as well as standard deviation (std) for Office and Office-Home. We observe that the std values are generally close to zero for all transfer tasks, which demonstrates the stability of our method.