While metric learning is important for Person re-identification (RE-ID), asignificant problem in visual surveillance for cross-view pedestrian matching,existing metric models for RE-ID are mostly based on supervised learning thatrequires quantities of labeled samples in all pairs of camera views fortraining. However, this limits their scalabilities to realistic applications,in which a large amount of data over multiple disjoint camera views isavailable but not labelled. To overcome the problem, we propose unsupervisedasymmetric metric learning for unsupervised RE-ID. Our model aims to learn anasymmetric metric, i.e., specific projection for each view, based on asymmetricclustering on cross-view person images. Our model finds a shared space whereview-specific bias is alleviated and thus better matching performance can beachieved. Extensive experiments have been conducted on a baseline and fivelarge-scale RE-ID datasets to demonstrate the effectiveness of the proposedmodel. Through the comparison, we show that our model works much more suitablefor unsupervised RE-ID compared to classical unsupervised metric learningmodels. We also compare with existing unsupervised RE-ID methods, and our modeloutperforms them with notable margins. Specifically, we report the results onlarge-scale unlabelled RE-ID dataset, which is important but unfortunately lessconcerned in literatures.
展开▼
