Deep part‐related feature learning for human‐level biometrics
Zhiheng Zhou, Zengqun Chen, Junchu Huang, Yusheng Zhang · IET Biometrics · 2019
Many human‐level biometrics systems rely on visual appearance to solve re‐identification problems. Most of them focus on recognizing local visual regions individually, without exploiting the relations between these separated visual regions during learning. In this study, the authors target on learning discriminative part‐related features and present novel network architecture, part relation network (PRN). It processes a set of part objects simultaneously in an end‐to‐end way to model relations of their appearance features, which strengthens the discriminative ability of local representation. Specially, the authors design PRN as a two‐branch deep network, each branch of which performs uniform partition with different numbers of stripes to obtain multi‐scale and cross‐over part features for identification. Experimental results demonstrate that this system achieves state‐of‐the‐art performance on three challenging data sets including Market‐1501, CUHK03, and DukeMTMC‐reID and demonstrate the effectiveness of modelling part relations for biometrics.