Introduction
Person Re-Identification (ReID) has seen significant advancements in daylight conditions, yet nighttime scenarios remain a formidable challenge. Surveillance at night is plagued by a "triple threat": low illumination, significant imaging noise, and unpredictable color shifts. Traditional approaches often rely on a sequential "enhance-then-recognize" pipeline, where an image is first brightened by an enhancement algorithm and then processed by a ReID model. However, this dependency often leads to suboptimal performance, as the ReID model becomes sensitive to artifacts or lost details in the reconstructed image.
The Collaborative Enhancement Network (CENet)
To overcome the limitations of sequential processing, this paper introduces the Collaborative Enhancement Network (CENet). Unlike previous methods, CENet adopts a parallel architecture that simultaneously handles image relighting and person identification.
Parallel Architecture and Feature Distillation
The core of CENet is a shared encoder that branches into two specialized subnets: a Relighting Subnet and a ReID Subnet. This design allows the ReID branch to leverage "lighting-aware" features from the relighting task without being strictly bound to the final visual quality of the reconstructed image.
The architecture utilizes a Vision Transformer (ViT) as a shared low-level feature extractor. High-level features are then processed through separate branches. A key innovation is the Feature Distillation module, which transfers semantic and lighting-enhanced knowledge from the relighting branch to the ReID branch. This ensures that the ReID model can "see" through darkness by utilizing the internal representations of the enhancement task.
Multi-domain Learning Strategy
Training deep models like ViT for nighttime ReID is difficult due to the scarcity of large-scale real-world datasets. The authors address this by introducing a Multi-domain Learning (MD) strategy and a new synthetic dataset, Syn_Dark.
Syn_Dark was created by applying multi-dimensional degradations—including lighting, contrast, and color shifts—to existing large datasets like Market1501 and MSMT17. During training, the model alternates between supervised learning on this large-scale synthetic data and unsupervised learning on real-world nighttime data (e.g., Night600). This hybrid approach ensures the model learns robust features from synthetic ground truths while generalizing effectively to the nuances of real surveillance footage.
Experimental Results
CENet demonstrates state-of-the-art performance across several benchmarks:
- Night600: Achieved a 25.00% Rank-1 accuracy and 13.30% mAP, significantly outperforming previous sequential methods.
- RGBNT201rgb: Reached 55.30% mAP, a substantial improvement over existing baselines.
- Efficiency: Since the relighting subnet is only required during training, the inference phase is streamlined, making CENet faster than methods requiring explicit image reconstruction.
Conclusion
CENet represents a shift from sequential to collaborative learning for nighttime ReID. By integrating relighting and identification in a parallel framework with feature distillation, the model effectively mitigates the impact of poor illumination. While challenges like high-frequency imaging noise persist, the combination of parallel processing and multi-domain learning sets a new standard for robust person re-identification in low-light environments. Future work will likely focus on reducing reliance on synthetic data through advanced unsupervised learning techniques.
Summary updated