Person Re-Identification via Unsupervised Transfer of Learned Visual Representations
Niki Martinel, Matteo Dunnhofer, Gian Luca Foresti, Christian Micheloni · 2017
Person re-identification is an open and challenging problem in computer vision. Most of the existing literature has focused on designing solutions which advantage of a supervised phase either to learn an end-to-end solution based on deep learning architectures or to find the optimal metric between hand-crafted features. Such solutions require a significant manual labor to annotate a large number of matching pairs to obtain good generalization performance. As a result, an extremely poor scalability is achieved. To address such a problem, we propose a deep learning scheme which leverages on the large quantity of labeled data that may be available in a source domain - different from the re-identification one- to learn a robust visual representation. This is then exploited in an unsupervised transfer learning scheme to better handle the difficulties in the target re-identification domain. The transferred sparse representation obtained via dictionary learning is used to perform the re-identification. Results on two benchmark datasets have shown that our approach performs on par or even better than state-of-the-art approaches.