An investigation over the state-of-the-art unsupervised domain adaptation person re-identification

Abstract

Over the years, person re-identification (re-ID) has been a crucial topic studied by many papers. In real word, person re-ID is usually used in the surveillance field and to recognise the exact person when given a photo over that person. And there are some existing pretrained models for the commonly used re-ID datasets like MSMT-17, DukeMTMC, Market-1501, and private models over the surveillance person datasets.

Given the situation, if we can utilise the existing models, it is convenient and desired to apply the original model to a bunch of new photos to generate a person re-ID model over the new dataset, since the relabelling process over a new dataset would be costly and it is hard for manpower to do the recognition. However, the question is that when this trained model applies directly on a different dataset, the recognition accuracy would significantly drop due to domain discrepancy between source model and the target dataset. And this is the major issue faced by the cross-domain person re-ID.

Hence, it is vital to provide a framework to utilise an existing model from labelled photos to the target unlabelled new photos to create a new re-ID model, which is usually called unsupervised domain adaptation method. The proposed solution aims to minimise the domain discrepancy and level up the confidence level of reusing an existing model.

In this paper, there would be a literature review on the unsupervised domain adaptation person re-ID (UDA re-ID), which mainly considers the summarization of the various types of state-of-the-art unsupervised domain adaptive re-ID method over past year studies. Different methods over the UDA re-ID would be discussed and compared and related training details and illustration figures would be provided.