Visible-Infrared Person Re-identification via Modality Augmentation and Center Constraints

Abstract

Visible-infrared person re-identification (VI-ReID) is a tricky cross-modality retrieval problem, where main challenge is the significant modality discrepancy. Existing methods mainly focus on bridging the relation between modalities by shared representation learning in the common embedding space. However, due to the outliers, these methods often struggle to build compact clustering subspaces. Besides, these methods also suffer from modality imbalance problem caused by more visible (VIS) modality images than infrared (IR) modality images, which induces the features of IR modality have to face an extra modality bias during the test phase. In this paper, we propose a novel modality augmentation and center constraints (MACC) framework, where grayscale channel joint random patch is adopted to generate semantic-rich images that are more similar to IR modality images in style, which reduces the modality discrepancy and, to some extent, alleviates the modality imbalance problem. In addition, based on the idea of partition, we design a fine-grained feature mining module (FFMM) to mine nuanced but discriminative information within each part, which is benefit to further alleviate the modality discrepancy. Meanwhile, to encourage more compact clustering, we propose a novel compact center clustering (C$^3$) loss to achieve intra-modality and inter-modality compact clustering. Apart from C$^3$ loss, we also introduce the center distribution consistencies (CDC) loss to align the feature vectors and prediction logits by optimizing the standard deviation and JS divergence, respectively. Extensive experiments on two public datasets demonstrate that our method achieves new state-of-the-art performance. The code will be available at https://github.com/cq0907/MACC.

Supported by organization Southwest University.