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

Xiao Teng,Long Lan,Dingyao Chen,Kele Xu,Nan Yin

2024-12-16

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: <a class="link-external link-https" href="https://github.com/tengxiao14/Neighbor-guided-USL-VI-ReID" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the common label noise problem in **Unsupervised Visible - Infrared Cross - Modal Person Re - identification (USL - VI - ReID)**. Specifically, existing unsupervised methods will encounter label noise during the learning process due to sub - optimal clustering results and significant modal differences, which hinders effective training.

### Problem Background

Unsupervised Visible - Infrared Cross - Modal Person Re - identification (USL - VI - ReID) aims to use unlabeled datasets to learn shared feature representations, thereby achieving cross - modal identity recognition. However, existing methods usually generate pseudo - labels through clustering. Due to the unsatisfactory clustering results and the significant differences between modalities, these pseudo - labels often contain noise, which affects the training effect of the model.

### Method Proposed in the Paper

To solve the above problems, the paper proposes a framework for alleviating common label noise based on neighbor information. Specifically, it includes two key modules:

1. **Neighborhood - Guided Universal Label Calibration Module (N - ULC)**:
- This module reduces label noise by introducing soft labels to replace hard pseudo - labels.
- For each sample, calculate its correlation with neighbor samples and generate more accurate soft labels based on these correlations.
- The formulas are as follows:
\[
\tilde{P}_{q_v}^{\text{intra}}=\left[\frac{|N(q_v, U_v, k)\cap C_v^l|}{|N(q_v, U_v, k)\cup C_v^l|}\right]
\]
\[
\tilde{I}_{q_v}^{\text{intra}}=\mu I_{q_v}^{\text{intra}}+(1 - \mu)P_{q_v}^{\text{intra}}
\]

2. **Neighborhood - Guided Dynamic Weighting Module (N - DW)**:
- This module reduces the influence of unreliable samples by using the consistency of neighbors and enhances the stability of training.
- According to the correlation between the sample and its neighbors, assign weights to each sample, thereby reducing the influence of unreliable samples on training.
- The formula is as follows:
\[
\omega_{q_v}^{\text{intra}}=\exp\left( - w\cdot(1 - [P_{q_v}^{\text{intra}}]_{l'})^2\right)
\]

### Experimental Results

The paper conducted extensive experiments on two public datasets (RegDB and SYSU - MM01). The results show that the proposed method outperforms existing methods in multiple settings. Especially in dealing with label noise, this method performs excellently and effectively improves the performance of the model.

### Summary

The main contributions of the paper are:
- Proposing a simple and effective framework that uses neighbor information to alleviate the common label noise in unsupervised Visible - Infrared Cross - Modal Person Re - identification.
- Introducing the Neighborhood - Guided Universal Label Calibration Module (N - ULC) and the Neighborhood - Guided Dynamic Weighting Module (N - DW), which are respectively used to improve the accuracy of identity representation and enhance the stability of training.
- The experimental results verify the effectiveness and superiority of this method.

Through these improvements, the paper solves the common label noise problem in existing methods and provides a new solution for unsupervised Visible - Infrared Cross - Modal Person Re - identification.