Relieving Universal Label Noise for Unsupervised Visible-Infrared Person Re-Identification by Inferring from Neighbors

TL;DR

This work tackles unsupervised visible–infrared person re-identification (USL-VI-ReID) by addressing universal label noise arising from suboptimal clustering and modality gaps. It introduces two neighbor-based modules: Neighbor-guided Universal Label Calibration (N-ULC) that converts hard pseudo labels into neighbor-informed soft labels, and Neighbor-guided Dynamic Weighting (N-DW) that down-weights unreliable samples, both integrated into a Progressive Graph Matching framework. A theoretical analysis based on a Rademacher generalization bound supports the approach, showing soft-label risk can be controlled as data size grows. Empirical results on RegDB and SYSU-MM01 demonstrate that the proposed, simple approach achieves state-of-the-art performance among USL-VI-ReID methods, validating the practical impact of leveraging neighbor information for label refinement and sample weighting.

Abstract

Unsupervised visible-infrared person re-identification (USL-VI-ReID) is of great research and practical significance yet remains challenging due to the absence of annotations. Existing approaches aim to learn modality-invariant representations in an unsupervised setting. However, these methods often encounter label noise within and across modalities due to suboptimal clustering results and considerable modality discrepancies, which impedes effective training. To address these challenges, we propose a straightforward yet effective solution for USL-VI-ReID by mitigating universal label noise using neighbor information. Specifically, we introduce the Neighbor-guided Universal Label Calibration (N-ULC) module, which replaces explicit hard pseudo labels in both homogeneous and heterogeneous spaces with soft labels derived from neighboring samples to reduce label noise. Additionally, we present the Neighbor-guided Dynamic Weighting (N-DW) module to enhance training stability by minimizing the influence of unreliable samples. Extensive experiments on the RegDB and SYSU-MM01 datasets demonstrate that our method outperforms existing USL-VI-ReID approaches, despite its simplicity. The source code is available at: https://github.com/tengxiao14/Neighbor-guided-USL-VI-ReID.

Figures (5)

• Figure 1: Illustration of the motivation of our method. Due to inferior clustering results, images of the same person can be divided into different groups in each modality. As a result, one-hot pseudo labels are limited to represent their identities in both homogeneous and heterogeneous learning processes. Additionally, the reliability of labeling can vary for diverse samples during the learning process.
• Figure 2: Framework of the proposed method. Based on the Progressive Graph Matching (PGM) framework, we propose the neighbor-guided universal label calibration module (Sec. \ref{['calibration']}) and neighbor-guided dynamic weighting module (Sec. \ref{['weighting']}), these module are applied on both the (b) homogeneous learning and (c) heterogeneous learning processes.
• Figure 3: Impact of hyper-parameter $k$ on different datasets.
• Figure 4: T-SNE visualization of features learned by PGM and our method on a subset of RegDB and SYSU-MM01 datasets. Different colors represent different identities.
• Figure 5: Impact of hyper-parameters $\lambda$ and $\mu$ on SYSU-MM01 dataset.