Dynamic Re-Weighting and Cross-Camera Learning for Unsupervised Person Re-Identification

Person Re-Identification (ReID) has witnessed tremendous improvements with the help of deep convolutional neural networks (CNN). Nevertheless, because different fields have their characteristics, most existing methods encounter problem poor generalization ability to invisible people. To address this problem, based on relationship between temporal and camera position, we propose a robust effective training strategy named smoothing dynamic re-weighting cross-camera learning (TSDRC). It uses algorithms transfer valuable knowledge labeled source domains unlabeled target domains. In domain stage, TSDRC iteratively clusters samples into several centers dynamically re-weights from each center score. Then, triplet loss is proposed fine-tune model. Particularly, improve discernibility CNN models in domain, generally shared person attributes margin-based softmax are adapted train terms clustered re-weighted center. Comprehensive experiments Market-1501 DukeMTMC-reID datasets demonstrate that method vastly improves performance unsupervised adaptability.