An enhanced siamese angular softmax network with dual joint-attention for person re-identification

Abstract

For person re-identification (re-ID), a core problem is how to learn discriminative feature representations of pedestrians. In this paper, we propose a novel enhanced siamese angular softmax network (ES-ASnet) to integrate identification, verification and metric learning into a unified network. First, a dual joint-attention (DJA) based identification model is proposed that can focus on both key local information and global contextual dependencies in spatial and channel domains simultaneously. Then, we adopt angular softmax (A-Softmax) loss in the training phase, which directly integrates metric learning into classification to enhance the discriminative capability of features in the angular space. Furthermore, the alignment module in the unified network can reduce the impact of misalignment between image pairs, which can further learn robust discriminative feature representations effectively. Experiments on three main person re-ID datasets, including Market1501, DukeMTMC-reID and CUHK03-NP, demonstrate that the proposed network has achieved competitive performance compared with several state-of-the-art methods for person re-ID.