Dual-level Modality Debiasing Learning for Unsupervised Visible-Infrared Person Re-Identification

TL;DR

The paper tackles modality bias in unsupervised visible-infrared person re-identification by introducing Dual-level Modality Debiasing Learning (DMDL). It unifies a causal-modeling module (CAI) that uses backdoor adjustment to isolate causal image-to-label relationships with a bias-free optimization strategy (CBT) that employs modality-specific data augmentation, label refinement, and feature alignment. Together, CAI and CBT enable learning of modality-invariant features and improve generalization across cross-modality scenarios, demonstrated on SYSU-MM01, RegDB, and LLCM with strong ablations. The approach delivers competitive unsupervised performance and highlights the value of combining causal interventions with bias-aware optimization in cross-modality visual tasks.

Abstract

Two-stage learning pipeline has achieved promising results in unsupervised visible-infrared person re-identification (USL-VI-ReID). It first performs single-modality learning and then operates cross-modality learning to tackle the modality discrepancy. Although promising, this pipeline inevitably introduces modality bias: modality-specific cues learned in the single-modality training naturally propagate into the following cross-modality learning, impairing identity discrimination and generalization. To address this issue, we propose a Dual-level Modality Debiasing Learning (DMDL) framework that implements debiasing at both the model and optimization levels. At the model level, we propose a Causality-inspired Adjustment Intervention (CAI) module that replaces likelihood-based modeling with causal modeling, preventing modality-induced spurious patterns from being introduced, leading to a low-biased model. At the optimization level, a Collaborative Bias-free Training (CBT) strategy is introduced to interrupt the propagation of modality bias across data, labels, and features by integrating modality-specific augmentation, label refinement, and feature alignment. Extensive experiments on benchmark datasets demonstrate that DMDL could enable modality-invariant feature learning and a more generalized model.

Figures (11)

• Figure 1: Existing USL-VI-ReID methods suffer from modality bias, leading to modality-related features. In contrast, our approach achieves modality-invariant feature learning through causal modeling and unbiased optimization. Green, yellow, and blue circles represent visible-specific, infrared-specific, and modality-shared information, respectively.
• Figure 2: The framework of the proposed DMDL. After obtaining cross-modality pseudo-labels through Iterative Maximum Confidence Alignment, the Causality-inspired Adjustment Intervention module is implemented for causal modeling to construct a low-biased model. Then, the Collaborative Bias-free Training strategy combines label refinement and modality alignment with data augmentation to optimize the model, further eliminating modality bias during training.
• Figure 3: (a) The structural causal model in cross-modality learning for USL-VI-ReID. (b) The modified structural causal model after the causal intervention.
• Figure 4: Illustration of the modality-specific augmentation. Circles represent channels of images. Subscript represents the sample index of pseudo-color images. For example, $R_2$, $G_2$, and $B_2$ are the red, green, and blue channels from the same pseudo-color image with index 2. The grey circle with ${IR}$ indicates the single channel of the infrared image.
• Figure 5: Detailed analysis of CBT on the SYSU-MM01 dataset under (a) all-search and (b) indoor-search modes. Rank-1 accuracy (%) and mAP (%) are reported.