Abstract
In a conventional domain adaptation person Re-identification (Re-ID) task, both the training and test images in the target domain are collected under sunny weather. However, in reality, pedestrians may be captured under severe weather conditions such as haze, dust, or snow. This paper proposes a novel Interference Suppression Model (ISM) to deal with the interference caused by hazy weather in domain adaptation person Re-ID. A teacher-student model is used in the ISM to distill interference information at the feature level by reducing the discrepancy between clear and hazy intrinsic similarity matrices. Furthermore, at the distribution level, an extra discriminator is introduced to assist the student model in making the interference feature distribution clearer. Experimental results show that the proposed method achieves superior performance on two synthetic datasets compared to state-of-the-art methods.
Introduction
Person Re-identification (Re-ID) is a critical task in surveillance, involving the matching of individuals across non-overlapping camera views. While Unsupervised Domain Adaptation (UDA) has significantly improved Re-ID performance by bridging the gap between labeled source and unlabeled target domains, most existing methods assume clear weather conditions. In real-world scenarios, inclement weather like haze introduces a significant domain shift that degrades the performance of standard models.
As illustrated in Figure 1, conventional models trained on clear data struggle when the target domain is obscured by haze. The haze not only blurs visual details but also fundamentally alters the data distribution, leading to incorrect rank-lists and a failure in robust identification.
Methodology: Interference Suppression Model (ISM)
The proposed ISM framework addresses environmental interference through two primary mechanisms: feature-level distillation and distribution-level alignment.
Teacher-Student Framework & Intrinsic Similarity Learning (ISL)
The ISM employs a Teacher-Student architecture where a Teacher model, pre-trained on clear images, guides a Student model processing hazy images. A key innovation is the use of Intrinsic Similarity Matrices. Rather than comparing raw features directly, the model focuses on the relative similarities between different people within a batch. The core assumption is that identity-based relationships (who looks like whom) should remain invariant to weather conditions. By minimizing the discrepancy between the Teacher's "clear" similarity matrix and the Student's "hazy" matrix, the Student learns to extract identity-relevant features while suppressing haze interference.
Adversarial Interference Distillation
To further refine the feature space, a discriminator is introduced to distinguish between features from clear and hazy images. Through adversarial training, the Student model is forced to generate "clean" features that are indistinguishable from clear-image features. Additionally, KL-divergence loss is utilized to align the Student's feature distribution with the Teacher's, ensuring consistency across domains.
Experiments and Results
Since standard Re-ID datasets lack hazy conditions, the authors synthesized Hazy-Market1501 and Hazy-DukeMTMC-reID using the Koschmieder physics-based model and monocular depth estimation.
The ISM model demonstrates substantial improvements over state-of-the-art UDA methods:
- Duke → Hazy-Market: ISM achieved a Rank-1 accuracy of 48.4% and an mAP of 21.9%, more than doubling the mAP of previous best methods like MMT.
- Market1501 → Hazy-Duke: ISM reached 37.7% Rank-1, significantly outperforming baseline competitors.
Ablation studies confirmed that both Intrinsic Similarity Learning and Adversarial Distillation are vital components; removing either significantly reduced performance.
Conclusion
The Interference Suppression Model (ISM) effectively mitigates the impact of hazy weather on person Re-ID by leveraging a teacher-student framework and adversarial learning. By focusing on structural relationships between images rather than raw pixel values, the model achieves robust performance in challenging environmental conditions. Future work may extend this framework to other inclement weather types such as rain or snow.
Summary updated