RGB-Infrared person re-IDentification (re-ID) aims to match RGB and infrared (IR) images of the same person. However, the modality discrepancy between RGB and IR images poses a significant challenge for re-ID. To address this issue, this paper proposes a Proxy-based Embedding Alignment (PEA) method to align the RGB and IR modalities in the embedding space. PEA introduces modality-specific identity proxies and leverages the sample-to-proxy relations to learn the model. Specifically, PEA focuses on three types of alignments: intra-modality alignment, inter-modality alignment, and cycle alignment. Intra-modality alignment aims to align sample features and proxies of the same identity within a modality. Inter-modality alignment aims to align sample features and proxies of the same identity across different modalities. Cycle alignment requires that a proxy is aligned with itself after tracing it along a cross-modality cycle (e.g., IR→RGB→IR). By integrating these alignments into the training process, PEA effectively mitigates the impact of modality discrepancy and learns discriminative features across modalities. We conduct extensive experiments on several RGB-IR re-ID datasets, and the results show that PEA outperforms current state-of-the-art methods. Notably, on SYSU-MM01 dataset, PEA achieves 71.0% mAP under the multi-shot setting of the indoor-search protocol, surpassing the best-performing method by 7.2%.

This paper explores the Vision Transformer (ViT) backbone for Unsupervised Domain Adaptive (UDA) person Re-Identification (Re-ID). While some recent studies have validated ViT for supervised Re-ID, no study has yet to use ViT for UDA Re-ID. We observe that the ViT structure provides a unique advantage for UDA Re-ID, i.e., it has a prompt (the learnable class token) at its bottom layer, that can be used to efficiently condition the deep model for the underlying domain. To utilize this advantage, we propose a novel two-stage UDA pipeline named Prompting And Tuning (PAT) which consists of a prompt learning stage and a subsequent fine-tuning stage. In the first stage, PAT roughly adapts the model from source to target domain by learning the prompts for two domains, while in the second stage, PAT fine-tunes the entire backbone for further adaption to increase the accuracy. Although these two stages both adopt the pseudo labels for training, we show that they have different data preferences. With these two preferences, prompt learning and fine-tuning integrated well with each other and jointly facilitated a competitive PAT method for UDA Re-ID.