Set-Based Feature Learning for Person Re-identification via Third-Party Images

Abstract

Person re-identification from disjoint camera views has been an important and unsolved problem due to large variations in illumination, viewpoint and pose. One way to attack this is by designing a new, more powerful image representation. However, we believe that existing representations are already sufficient. The main difficulty is how to pick the most informative information using these representations. Inspired by the prototype theory from the cognition field and Exemplar-SVM, we propose a novel and simple set-based feature learning re-identification method via third-party images. In our settings, each query/gallery example is an image set of the same individual, not just a single image. Discriminative features of a certain individual image set are explored from the third-party images. Comparisons with state-of-the-art methods on benchmark datasets demonstrate impressive results using simple and common features.