Abstract
Person re-identification (re-id) aims to retrieve the image which belongs to the same person from a set of images according to a few clues. The dataset images are usually from multiple cameras of different locations. Recently, local feature extraction has been one of the mainstream to handle this task. Recent methods show that exploration of part-level characteristics in person re-id can be helpful for its ability in capturing discriminative features. However, those works ignore the association among image regions, and it will lead to wrong retrieval results when not considering such clues. To address this limitation, we propose an effective Global Correlative Network (GCN) to jointly exploit the potential relationships between the features from semantic regions for more discriminative image representations. Specifically, we integrate the association between local areas and the global for discriminative region-aggregated features. With a relatively concise structure, a novel relation learning framework is proposed to enhance the representation with the contextual relations between the local and global perspectives. Experiments on Market-1501, DukeMTMC-reID, and CUHK03 show the effectiveness of our proposed method. Specifically, our method achieves a state-of-the-art rank-1 accuracy of 90.0%, 83.7%, 78.5% on the DukeMTMC-reID, CUHK03(Labeled), and CUHK03(Detected) datasets, so it sets a new state-of-the-art.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have