Towards multi-scale deep features learning with correlation metric for person re-identification

Previous person re-identification (Re-ID) methods usually focus on extracted features to against the appearance variations of pedestrians under different circumstances. The problem caused by the scale variations did not attract much attention and is not well addressed either. In this work, we propose a novel Multi-scale Deep Feature Learning with correlation metric (MDFLCM) model to handle the scale problem in Re-ID. Specifically, multi-scale high-level features are extracted by a specially designed end-to-end multi-scale deep convolutional network (MS-DCN) at various resolution levels. By adding an extra correlation layer in our MDFLCM model, we can achieve the accuracy of image patch matching up to pixel-wise level. Different from other methods extracting multi-scale features through multiple networks, we extract multi-scale features via a single network with one input image. Extensive comparative evaluations with state-of-the-art methods on four public datasets: CUHK01, CUHK03, Market 1501 and DukeMTMC-reID, demonstrate the effectiveness of the proposed MDFLCM model on Re-ID.