Unsupervised Person Re-identification via Discriminative Exemplar-level and Patch-level Feature Fusion

The majority of existing person re-identification(re-ID) approaches adopt supervised learning pattern, which require large amount of labeled data to train models. However, due to the high cost of marking by hand, they are limited to be widely used in reality. On the other hand, due to the difference of the camera angle, there are many variations in pedestrian postures and illumination. It is known that Extracting discriminative features is pretty effective to solve the problem of person re-ID. Therefore, we propose to fuse exemplar-level features and patch-level features to obtain more distinguishing pedestrian image features for unsupervised person re-ID. Firstly, We carefully design exemplar-level and patch-level feature learning framework(EPFL). The skeleton frame adopts bicomponent branch, one branch is used to learn the global features of pedestrian images, the other is used to learn local features. Then, the global features at the example level and local features at the patch level are fused, thus the discriminative pedestrian image features can be obtained. Furthermore, feature memory bank (FMB) is introduced to facilitate the calculation of the similarity between pedestrian images on unlabeled dataset. We carry on our proposed method on two frequently-used datasets, namely, Market-1501 and DukeMTMC-reID dateset. Experimental results clearly demonstrate the advantage of the proposed approach for unsupervised person re-ID.