Disentangling style on dynamic aligned poses for individual identification

Abstract Despite the rapid increase of research on individual identification in recent years, most of them focused on extracting the invariant visual cues of individuals. However, few efforts have been devoted to exploring the inherent feature caused by physiological differences. The main challenge in this task arises from two aspects: (i) the individual inherent feature (style) of each pose may be unique; (ii) the individual inherent feature and the shared feature from pose are always coupled together. In this paper, we propose a novel model, namely Disentangling Style Network (DS-Net), which contains an alignment module to match similar poses from different individuals, and a disentangling module to encode the individual inherent feature and shared feature separately. Theoretically, the sub-modules in our model are mutually beneficial for each other, the alignment module can improve the quality of training poses pairs in the disentangling module, meanwhile, the shared feature encoded by the disentangling module can also improve the accuracy of the alignment module. The coupled network contains a set of functional components, can be efficiently trained by gradient-based optimizers in a practical iterative way. We evaluate the proposed DS-Net on several public databases, matching or outperforming the state-of-the-art approaches. The code is publicly available 1 .