Hazy Re-ID: An Interference Suppression Model for Domain Adaptation Person Re-Identification Under Inclement Weather Condition

Jian Pang, Dacheng Zhang, Huafeng Li, Weifeng Liu, Zhengtao Yu

TL;DR

Researchers have developed an Interference Suppression Model (ISM) to solve the problem of identifying people across camera networks in hazy or inclement weather. By using a teacher-student framework and adversarial learning, the model "distills" away the interference caused by haze, significantly outperforming existing domain adaptation methods on newly created synthetic hazy datasets.

What Problem Does This Paper Address?

Person Re-identification (Re-ID) is the task of spotting the same individual across different, non-overlapping camera views. While modern AI models are getting quite good at this, they have a major "fair weather" bias. Most models are trained and tested on datasets collected on clear, sunny days.

In the real world, surveillance systems must operate in fog, dust, and snow. Haze doesn't just make images look blurry; it fundamentally alters the data distribution, creating a "domain gap" that causes standard models to fail. Existing Unsupervised Domain Adaptation (UDA) methods—which try to adapt a model from a labeled "source" domain to an unlabeled "target" domain—don't account for this environmental interference. This paper tackles the specific challenge of performing Re-ID when the target environment is hazy.

Methods

The authors propose the Interference Suppression Model (ISM), which relies on two main strategies to "clean" the features extracted from hazy images:

  1. Teacher-Student Framework & Intrinsic Similarity Learning (ISL): The system uses a "Teacher" model (trained on clear images) to guide a "Student" model (which sees hazy images). Instead of just comparing raw features, they use Intrinsic Similarity Matrices. The idea is that the relationship between different people in a batch (who looks like whom) should remain the same whether the weather is clear or hazy. By minimizing the difference between the Teacher’s clear similarity matrix and the Student’s hazy one, the Student learns to ignore the haze.

  2. Interference Distillation via Adversarial Learning: To further refine the features, the authors introduce a Discriminator. This component tries to determine if a feature set came from a clear image or a hazy one. Through adversarial training, the Student model learns to produce features that are so "clean" the Discriminator can't tell they originated from a hazy image. They also use KL-divergence loss to ensure the Student’s feature distribution matches the Teacher’s "clear" distribution.

Datasets

Since standard Re-ID datasets are mostly sunny, the authors created two new benchmarks by applying a physical haze model (the Koschmieder model) and monocular depth estimation to existing data:

  • Hazy-Market1501: Based on the popular Market-1501 dataset, containing over 32,000 generated hazy images.
  • Hazy-DukeMTMC-reID: Based on the DukeMTMC-reID dataset, adapted for hazy conditions.

These datasets allow for testing how well a model can adapt from a clear source (e.g., Duke) to a hazy target (e.g., Hazy-Market).

Key Insights and Ideas

  • Haze as a Domain Gap: The paper frames inclement weather not just as "noise," but as a distinct domain shift. This allows them to leverage UDA techniques rather than just simple image de-hazing.
  • Relationships over Raw Pixels: The most clever "aha moment" is the use of the Intrinsic Similarity Matrix. By focusing on how images relate to each other in a batch, the model finds a structural anchor that is more robust to environmental interference than individual pixel values.
  • Attention Matters: Visualization (using Grad-CAM) shows that standard models get "distracted" by haze, focusing on the background or random noise. The ISM model successfully forces the network's "attention" back onto the person’s body and identifying features.

Key Results

The ISM model shows a dramatic improvement over state-of-the-art (SOTA) domain adaptation methods like ECN and MMT:

  • Duke → Hazy-Market: The ISM achieved a Rank-1 accuracy of 48.4% and an mAP (mean Average Precision) of 21.9%. For context, the previous best method (MMT-500) only hit 29.9% Rank-1 and 12.4% mAP.
  • Market1501 → Hazy-Duke: ISM reached 37.7% Rank-1, significantly outperforming the baseline and other UDA competitors.
  • Ablation Studies: The researchers proved that both the Similarity Learning (ISL) and the Adversarial Distillation were necessary; removing either one led to a significant drop in performance.

Limitations and Future Work

  • Synthetic vs. Real Haze: The primary limitation is the reliance on synthetic hazy datasets. While the physics-based model used is sophisticated, real-world haze can be non-uniform and vary in chemical composition, which might affect light scattering differently.
  • Generalization to Other Weather: While the paper focuses on haze, the authors suggest that the "ISM" framework could be generalized. Future work could involve testing the model against rain, snow, or even low-light "nighttime" Re-ID, which presents its own set of unique interference patterns.

Summary updated

Abstract

In a conventional domain adaptation person Re-identification (Re-ID) task, both the training and test images in target domain are collected under the sunny weather. However, in reality, the pedestrians to be retrieved may be obtained under severe weather conditions such as hazy, dusty and snowing, etc. This paper proposes a novel Interference Suppression Model (ISM) to deal with the interference caused by the hazy weather in domain adaptation person Re-ID. A teacher-student model is used in the ISM to distill the interference information at the feature level by reducing the discrepancy between the clear and the hazy intrinsic similarity matrix. Furthermore, in the distribution level, the extra discriminator is introduced to assist the student model make the interference feature distribution more clear. The experimental results show that the proposed method achieves the superior performance on two synthetic datasets than the state-of-the-art methods. The related code will be released online https://github.com/pangjian123/ISM-ReID.
BibTeX
@inproceedings{pang2021hazy,
  author = {Jian Pang and Dacheng Zhang and Huafeng Li and Weifeng Liu and Zhengtao Yu},
  title = {{Hazy Re-ID: An Interference Suppression Model for Domain Adaptation Person Re-Identification Under Inclement Weather Condition}},
  booktitle = {ICME},
  year = {2021},
  url = {https://openreview.net/forum?id=-a7wKo-fTLj},
}

Loading related content...

Figure