Fast and Accurate Person Re-identification with Xception Conv-Net and C2F.

Person re-identification (re-id) is the task of identifying a person of interest across disjoint camera views in a multi-camera system. This is a challenging problem due to the different poses, viewpoints and lighting conditions. Deeply learned systems have become prevalent in the person re-identification field as they are capable to deal with the these obstacles. Conv-Net using a coarse-to-fine search framework (Conv-Net+C2F) is such a deeply learned system, which has been developed with both a high-retrieval accuracy as a fast query time in mind. We propose three contributions to improve Conv-Net+C2F: (1) training with an improved optimizer, (2) constructing Conv-Net using a different Convolutional Neural Network (CNN) not yet used for person re-id and (3) coarse descriptors having fewer dimensions for improved speed as well as increased accuracy. With these adaptations Xception Conv-Net+C2F achieves state-of-the-art results on Market-1501 (single-query, 72.4% mAP) and the new, challenging data split of CUHK03 (detected, 42.6% mAP).