Unsupervised cross-domain person re-identification by instance and distribution alignment

• A novel idea of exploring instance-wise localised source knowledge for unsupervised cross-domain person re-id. Hierarchical Unsupervised Domain Adaptation method designed to discover at the instance level. Analyse feature representations domain adaptation in closed-set supervised learning vs. open-set learning. Most existing re-identification (re-id) methods assume model training on a separate large set samples from target domain. While performing well domain, such trained models are seldom generalisable new independent without further labelled data To solve this scalability limitation, we develop (HUDA) method. It can transfer information an dataset (a domain) unlabelled Specifically, HUDA is jointly global distribution alignment and local two-level hierarchy discovering transferable adaptation. Crucially, approach aims overcome under-constrained problem methods. Extensive evaluations show superiority re-id over wide variety state-of-the-art four benchmarks: Market-1501, DukeMTMC, MSMT17 CUHK03.