- Conference Article
- 10.1109/aibdf67964.2025.11440878
An Adaptive Dual-domain Feature Alignment Network for All-Day City Image Analysis
- Dec 26, 2025
- Ying Qian + 3 more +3
The vision of smart cities relies on robust, all-day automated image analysis systems to enhance urban management and public security. However, severe illumination variations between day and night present a significant challenge for tasks such as vehicle re-identification, where the same vehicle can appear drastically different due to lighting changes and nighttime glare. This impedes the reliability of continuous urban image analytics. To address this issue, we propose an Adaptive Dual-domain Feature Alignment Network (ADFA-Net). Our approach aligns vehicle features across time domains by learning illumination-invariant representations. ADFA-Net employs a dual-stream architecture with dedicated branches for processing day and night images. Each branch incorporates a glare suppression module to enhance feature quality. Crucially, features from both streams are adaptively fused at multiple semantic levels to construct a coherent and aligned feature space. Extensive evaluations on the cross-day-night benchmarks DN-348 and DN-Wild demonstrate the effectiveness of our model. On DN-348, ADFA-Net achieves 48.7% mAP and 72.2% Rank-1 accuracy. On the more challenging DN-Wild dataset, it attains 41.4% mAP and 51.0% Rank-1 accuracy.
Read more