Abstract
Person re-identification (Re-ID) has achieved significant success in daytime scenarios, yet nighttime performance remains a critical bottleneck due to extreme noise and low illumination. This paper introduces KnightReid, the first large-scale benchmark dataset specifically designed for nighttime RGB person Re-ID. To address the unique challenges of nighttime data, we propose a joint learning framework that integrates an image denoising subnet with a Re-ID feature extraction subnet. By optimizing these components simultaneously, our model learns to preserve identity-critical features while suppressing sensor noise, significantly outperforming standard daytime models on nighttime benchmarks.
The Challenge of Nighttime Surveillance
Person re-identification—the task of matching an individual across non-overlapping camera views—is essential for security and robotics. However, existing research suffers from a "daytime bias," as popular datasets like Market-1501 are captured in optimal lighting. In contrast, nighttime surveillance presents three major hurdles:
- Severe Sensor Noise: Low-light conditions produce grainy images that obscure fine details.
- Color Degradation: Nighttime images often lose the vibrant color information that models typically use for discrimination.
- Illumination Extremes: Scenes often feature a mix of deep shadows and intense glare from artificial light sources.
As shown in the figure above, the visual quality of nighttime probes and galleries is significantly degraded compared to daytime counterparts, making standard feature extraction unreliable. The authors argue that current Re-ID models fail in these conditions because they were not designed to handle the specific noising properties of nighttime data.
The KnightReid Benchmark
To facilitate research in this area, we present KnightReid, a large-scale dataset captured in a real-world campus environment.
- Scale: 937 unique identities across 3 high-definition cameras.
- Volume: Over 315,000 hand-labeled bounding boxes.
- Realism: Unlike datasets using infrared or cross-modality data, KnightReid focuses on real-world nighttime RGB footage, reflecting standard security infrastructure.
Proposed Methodology: Joint Denoising and Re-ID
We propose a "two-subnet" pipeline that treats denoising and identification as a unified task rather than two separate steps.
Denoising Subnet
We utilize a Non-local Recurrent Network (N3) to clean input images. Unlike generic filters, N3 leverages structural similarities within the image to recover high-quality details from noisy inputs.
Re-ID Subnet
For feature extraction, we employ the Harmonious Attention Convolutional Neural Network (HACNN). This architecture is particularly effective at focusing on informative body regions (e.g., torso, legs) while ignoring background noise and clutter.
Joint Training Strategy
The core innovation is the joint optimization of both subnets. Rather than treating denoising as a pre-processing step, we use a combined loss function: $$L_{total} = \lambda L_{psnr} + (1 - \lambda) L_{tri}$$ The $L_{psnr}$ term ensures effective denoising, while the Triplet Loss ($L_{tri}$) ensures discriminative power. This approach forces the denoising network to specifically preserve features that are most useful for identification.
Key Insights and Results
Our experiments yield several critical findings for nighttime Re-ID:
- Denoising as a Feature: Standard denoisers like BM3D provide marginal gains. The best performance is achieved when the denoising layer is optimized specifically to support the identification layer.
- Robustness of Texture: While color-based features (like LOMO) fail at night, texture-based methods (like GOG) remain more robust as silhouettes and patterns are still partially visible.
- The Role of Attention: Attention modules are vital in the dark. By focusing on the person's silhouette, the model can effectively ignore the "snowy" noise prevalent in the background.
The proposed joint model consistently outperformed baselines, showing an improvement in mean Average Precision (mAP) of 0.5% to 1.1% and significant boosts in Rank-1 accuracy across various camera pairs.
Conclusion and Future Work
KnightReid provides a necessary foundation for robust 24/7 surveillance. While our joint learning framework shows promise, future work must address the computational costs of deep denoising for real-time applications and expand the dataset to include diverse weather conditions like rain or fog, which introduce different types of visual distortion.
Summary updated