Person Re-identification with Joint-Loss

Abstract

Person re-identification is a technique that search the given target in the video surveillance network. This technique has been widely applied to security and surveillance system, and also become a research hotspot in computer vision. Person re-identification has been challenging due to the large number of cameras in the network and variation in camera angles, illumination, occlusion and poses. In this paper, we proposed a person re-id approach that can resist occlusions and variations based on a human pose guided convolution neural network framework with joint loss functions. We extract local features from body parts localized by landmarks, merge it with global features to learn the similarity metric. Identification loss and pose-constrained triplet loss function are jointly employed to train the model. Our approach outperforms most state-of-the-art methods on three large-scale datasets, with an accuracy of 83.31%, 86.1% and 72.6% on Cuhk03, Market1501 and Duke MTMC-reID respectively.