Retinex-based feature-consistent domain adaptation for nighttime object detection

Nighttime object detection is highly challenging, primarily due to the lack of annotated data and the complex noise in low-light images. Many domain adaptation methods have been applied to nighttime object detection tasks and have achieved promising results. However, existing domain adaptation methods suffer from several limitations. On the one hand, these methods do not explicitly guide the domain adaptation network to learn inter-domain consistent features, resulting in limited domain adaptation capability. On the other hand, the unsupervised learning approaches based on pseudo-labels struggle to ensure the quality of pseudo-labels, which adversely affects the performance of the trained model. To address these issues, we propose a feature-consistent domain adaptation method for nighttime object detection based on Retinex decomposition and reconstruction with classification contrastive learning. Our method achieves state-of-the-art (SOTA) performance on the BDD100K and SHIFT datasets and outperforms supervised training methods directly applied to the target domain.