Open Set Domain Adaptation using Optimal Transport

We present a 2-step optimal transport approach that performs a mapping from a\nsource distribution to a target distribution. Here, the target has the\nparticularity to present new classes not present in the source domain. The\nfirst step of the approach aims at rejecting the samples issued from these new\nclasses using an optimal transport plan. The second step solves the target\n(class ratio) shift still as an optimal transport problem. We develop a dual\napproach to solve the optimization problem involved at each step and we prove\nthat our results outperform recent state-of-the-art performances. We further\napply the approach to the setting where the source and target distributions\npresent both a label-shift and an increasing covariate (features) shift to show\nits robustness.\n