Partial person re-identification using a pose-guided alignment network with mask learning

摘要

Partial person re-identification is a challenging task, in which only a partial observation of a person is available. There is severe misalignment when directly comparing a partial image with the holistic image, which leads to performance degradation with re-identification algorithms. In this paper, we propose a pose-guided alignment and mask learning network (PMN) to solve the problems of large parts missing and significant pedestrian misalignment. The proposed model includes a pose-guided spatial transformer (PST) module and a masked feature extractor. The PST module samples an affine transformed image from a holistic/partial image to align the pedestrian image with a standard pose. The masked feature extractor, which consists of a backbone network and a mask learning branch (MLB), is designed to learn the visibility of body parts to select effective features. The experimental results on two reported partial person benchmarks show that the proposed method achieves competitive performance compared to that of state-of-the-art methods.