Unsupervised Visible-Infrared ReID via Pseudo-label Correction and Modality-level Alignment

Yexin Liu, Weiming Zhang, Athanasios V. Vasilakos, Lin Wang

Abstract

Unsupervised visible-infrared person re-identification (UVI-ReID) has recently gained great attention due to its potential for enhancing human detection in diverse environments without labeling. Previous methods utilize intra-modality clustering and cross-modality feature matching to achieve UVI-ReID. However, there exist two challenges: 1) noisy pseudo labels might be generated in the clustering process, and 2) the cross-modality feature alignment via matching the marginal distribution of visible and infrared modalities may misalign the different identities from two modalities. In this paper, we first conduct a theoretic analysis where an interpretable generalization upper bound is introduced. Based on the analysis, we then propose a novel unsupervised cross-modality person re-identification framework (PRAISE). Specifically, to address the first challenge, we propose a pseudo-label correction strategy that utilizes a Beta Mixture Model to predict the probability of mis-clustering based network's memory effect and rectifies the correspondence by adding a perceptual term to contrastive learning. Next, we introduce a modality-level alignment strategy that generates paired visible-infrared latent features and reduces the modality gap by aligning the labeling function of visible and infrared features to learn identity discriminative and modality-invariant features. Experimental results on two benchmark datasets demonstrate that our method achieves state-of-the-art performance than the unsupervised visible-ReID methods.

Introduction

Person re-identification (ReID) is the task of recognizing the same individual across different cameras. While standard ReID uses visible light (RGB) cameras, these fail in low-light or nighttime conditions. Visible-Infrared ReID (VI-ReID) solves this by using infrared (IR) sensors, but it introduces a massive "modality gap"—the same person looks fundamentally different in thermal vs. color images.

Most existing methods rely on massive, manually labeled datasets. However, labeling infrared images is tedious and error-prone. Unsupervised VI-ReID (UVI-ReID) aims to learn these features without labels. The authors identify two critical hurdles:

  1. Noisy Pseudo-labels: Unsupervised methods usually group images into "clusters" to create fake labels. If the clustering is wrong, the model learns incorrect features.
  2. Identity Misalignment: Simply trying to make the overall distribution of RGB and IR features look similar (marginal alignment) often accidentally aligns the wrong people, confusing different identities across modalities.

The PRAISE Framework

The authors propose PRAISE (Pseudo-label Rectification and Alignment for Infrared-Svisible Environment), which consists of two main pillars to address the aforementioned challenges.

Overview of the PRAISE framework: (a) Pseudo-label Correction (PLC) handles noisy clusters via BMM; (b) Modality-level Alignment (MLA) bridges the RGB-IR gap using feature translation.

1. Pseudo-label Correction (PLC)

To handle noisy clusters, the authors leverage the "memory effect" of neural networks—the observation that models tend to learn simple, clean patterns before they start "memorizing" noisy outliers.

  • They use a Beta Mixture Model (BMM) to estimate the probability that a specific sample is "noisy" based on its loss value.
  • A perceptual term is added to the contrastive learning loss. If a sample is likely noisy, the model relies less on its assigned cluster label and more on its closest feature neighbors, preventing the noise from degrading the model.

2. Modality-level Alignment (MLA)

To bridge the RGB-IR gap without labels, the model needs to know which RGB cluster corresponds to which IR cluster.

  • Bi-directional Translation (BiT): Using a CycleGAN-style approach, the model generates "pseudo-infrared" features from real visible features (and vice versa).
  • Sequential Filtering Matching (SFM): Instead of guessing, the model uses the translated features to find the best match between real RGB clusters and real IR clusters. If translated RGB features consistently fall into a specific IR cluster, the model concludes those two clusters represent the same person.

Experimental Results

The model was evaluated on two standard benchmark datasets for cross-modality ReID: SYSU-MM01 and RegDB. The PRAISE framework significantly outperformed previous unsupervised methods:

  • SYSU-MM01 (All-search): Achieved a Rank-1 accuracy of 59.44% and 53.27% mAP (mean Average Precision). This represents nearly a 2% improvement in Rank-1 over the previous state-of-the-art (PGMAL).
  • RegDB: Achieved 72.54% Rank-1 accuracy (Visible to Thermal), showing strong performance even on smaller datasets.
  • Qualitative Success: t-SNE visualizations showed that after applying MLA, the clusters for visible and infrared features of the same identity moved significantly closer together, effectively "closing the gap."

Conclusion and Future Work

PRAISE addresses the "double-blind" challenge of UVI-ReID by combining a noise-aware pseudo-label correction strategy with a modality-translation alignment module. While the framework achieves state-of-the-art performance, the authors acknowledge the complexity of the current three-stage training process and aim to simplify it into a more end-to-end stable process in future iterations. Additionally, they suggest exploring "source-free" methods to enhance privacy by avoiding exposure of sensitive visible-light data during the adaptation phase.

Summary updated

Abstract

Unsupervised visible-infrared person re-identification (UVI-ReID) has recently gained great attention due to its potential for enhancing human detection in diverse environments without labeling. Previous methods utilize intra-modality clustering and cross-modality feature matching to achieve UVI-ReID. However, there exist two challenges: 1) noisy pseudo labels might be generated in the clustering process, and 2) the cross-modality feature alignment via matching the marginal distribution of visible and infrared modalities may misalign the different identities from two modalities. In this paper, we first conduct a theoretic analysis where an interpretable generalization upper bound is introduced. Based on the analysis, we then propose a novel unsupervised cross-modality person re-identification framework (PRAISE). Specifically, to address the first challenge, we propose a pseudo-label correction strategy that utilizes a Beta Mixture Model to predict the probability of mis-clustering based network's memory effect and rectifies the correspondence by adding a perceptual term to contrastive learning. Next, we introduce a modality-level alignment strategy that generates paired visible-infrared latent features and reduces the modality gap by aligning the labeling function of visible and infrared features to learn identity discriminative and modality-invariant features. Experimental results on two benchmark datasets demonstrate that our method achieves state-of-the-art performance than the unsupervised visible-ReID methods.
BibTeX
@article{liu2025unsupervised,
  author = {Yexin Liu and Weiming Zhang and Athanasios V. Vasilakos and Lin Wang},
  title = {{Unsupervised Visible-Infrared ReID via Pseudo-label Correction and Modality-level Alignment}},
  journal = {IEEE Transactions on Neural Networks and Learning Systems},
  year = {2025},
  doi = {10.1109/tnnls.2025.3591641},
}
Figure