Illumination Distillation Framework for Nighttime Person Re-Identification and A New Benchmark

Andong Lu, Zhang Zhang 0001, Yan Huang, Yifan Zhang 0004, Chenglong Li, Jin Tang, Liang Wang

Introduction

Person Re-Identification (Re-ID) is a critical task in computer vision, focused on matching individuals across non-overlapping camera views. While significant progress has been made in daytime scenarios, nighttime Re-ID remains a significant challenge due to poor imaging quality and a lack of diverse, real-world datasets. Low-light conditions introduce sensor noise and low contrast, making standard feature extraction unreliable. Furthermore, existing nighttime datasets often rely on synthetic darkening of daytime images, which fails to capture the physical complexities of actual nighttime photography.

The Illumination Distillation Framework (IDF)

To address these challenges, the authors propose the Illumination Distillation Framework (IDF). This architecture employs a dual-branch approach combined with a teacher-student learning strategy to extract robust features from low-light images.

The Illumination Distillation Framework (IDF) architecture, featuring the Master Branch (MBranch), Illumination Enhancement Branch (IEBranch), and the Illumination Distillation Module (IDModule).

The framework consists of three primary components:

  1. Master Branch (MBranch): Extracts features directly from the original nighttime image ($I_0$).
  2. Illumination Enhancement Branch (IEBranch): Utilizes a DCE-Net sub-network to dynamically enhance the brightness and contrast of the input image before feature extraction.
  3. Illumination Distillation Module (IDModule): Acts as the core fusion mechanism. It takes features from both branches and passes them through a bottleneck fusion model (encoder-decoder) to filter out noise and distill essential identity information.

During training, the IDModule serves as a "teacher," guiding the individual branches via Kullback-Leibler (KL) divergence. This distillation strategy ensures that both branches learn to prioritize discriminative features while ignoring lighting-induced artifacts.

The Night600 Dataset

A major contribution of this work is the introduction of Night600, a large-scale, real-world nighttime Re-ID benchmark. Unlike previous datasets that use synthetic darkness, Night600 contains 28,813 images of 600 identities captured by 8 non-overlapping cameras in diverse outdoor environments, including campus roads lit by street lamps and headlights. This dataset highlights the "performance cliff" where models trained on daytime or synthetic data fail significantly in real nighttime conditions.

Experimental Results

The IDF framework was evaluated against several state-of-the-art (SOTA) models. The results demonstrate that real-world nighttime Re-ID is substantially more difficult than daytime or synthetic tasks.

Rank-1 accuracy comparison of various Re-ID models on Night600, Market1501-N (synthetic nighttime), and Market1501 (daytime).

Key findings include:

  • Performance Gains: On the Night600 dataset, the IDF improved the Rank-1 accuracy of the baseline IDE+ model from 7.75% to 13.63%. Using a transformer backbone, it reached 17.20% Rank-1, surpassing advanced models like TransReID (16.00%) and AGW (12.50%).
  • Synthetic vs. Real Gap: As shown in the performance comparison, models that achieve over 90% Rank-1 on daytime (Market1501) or synthetic nighttime (Market1501-N) data drop to below 20% on the real-world Night600 dataset.
  • Ablation Studies: The combination of dual branches and the distillation module proved more effective than using either branch in isolation, confirming that enhancement and raw feature extraction are complementary.

Conclusion and Future Work

The Illumination Distillation Framework and the Night600 dataset represent a significant step forward for nighttime surveillance. By treating illumination enhancement as a source of complementary information rather than a direct replacement for raw data, the IDF effectively mitigates the noise inherent in low-light sensors.

Despite these gains, the low absolute accuracy (approx. 17% Rank-1) underscores the ongoing difficulty of the task. Future research will focus on reducing computational complexity for real-time deployment and improving cross-domain generalization to make nighttime Re-ID more robust across varying lighting environments.

Summary updated

BibTeX
@article{lu2024illumination,
  author = {Andong Lu and Zhang Zhang and Yan Huang and Yifan Zhang and Chenglong Li and Jin Tang and Liang Wang},
  title = {{Illumination Distillation Framework for Nighttime Person Re-Identification and A New Benchmark}},
  journal = {IEEE Trans. Multimedia},
  year = {2024},
  doi = {10.1109/tmm.2023.3266066},
}
Figure