Momentum source-proxy guided initialization for unsupervised domain adaptive person re-identification

Abstract

Unsupervised domain adaptive person re-identification (UDA Re-ID), aiming to adapt the model trained from source domain to target domain, is especially challenging due to the non-overlapping identities between the two Re-ID domains. State-of-the-art UDA Re-ID methods optimize the model pre-trained on source domain with pseudo labels generated by clustering algorithms on the target domain. The drawback lies in that the initial parameters are learned only from labeled source domain, neglecting the target domain information that can be easily obtained from unlabeled data. In order to better fit the target distribution while preventing from over-fitting to the source one, we propose a novel momentum source-proxy guided initialization (MSPGI) approach to integrate information from unlabeled data into the pre-training process. Specifically, we assign soft labels to unlabeled data according to similarity to the feature proxies of the source domain, based on the finding that different Re-ID datasets share commonalities. In addition, we instantiate the pretext task in unsupervised pre-training as constraining the predicted soft label to be consistent with the one calculated from the temporally-averaged parameters of the model. Experiments are conducted on multiple downstream approaches, pushing forward the state-of-the-art results by an impressive margin on Market-1501 and DukeMTMC-reID. By making use of unlabeled data, MSPGI further improves the performance of a fully supervised network.