Semi-supervised cross-modality person re-identification based on pseudo label learning

Abstract

Visible-infrared person re-identification (RGB-IR Re-ID) aims to find images of the same identity from different modalities. In practice, multiple person and cameras can provide abundant training samples and non-negligible modality differences makes manual labeling of all samples be impractical. How to accurately re-identify cross-modality pedestrians under the training condition of having few labeled samples and a quantity of unlabeled samples is an important research question. However, person re-identification in this scenario, which we call Semi-Supervised Cross-Modality Re-ID (SSCM Re-ID), has not been well studied. In this paper, we propose a cross-modality pseudo label learning (CPL) framework for SSCM Re-ID task. It consists of three modules: the feature mapping module, the identity alignment module and the pseudo-label generation module. The feature mapping module is designed to extract shared discriminatory features from modality-specific channels, followed by the identity alignment module that aims to align person identities jointly at the global-level and part-level aspects. Finally, the pseudo-label generation module is used to select samples with reliable pseudo labels from the unlabeled samples based on the confidence level. Moreover, we propose the dynamic center-based cross-entropy loss to constrain the distance of similar samples. Experiments on widely used cross-modality Re-ID datasets demonstrate that CPL can achieve the state-of-the-art SSCM Re-ID performance.