Abstract: Nighttime object detection faces numerous challenges compared to daytime scenarios, primarily due to low-light conditions and the scarcity of high-quality annotated data, which hinder feature extraction and degrade detection accuracy. To this end, a domain-adaptive object detection method for nighttime images is proposed. First, a nighttime domain-adapted luminance alignment module is designed to convert a labeled daytime source domain image into a corresponding nighttime target domain image. Bridging the gap between source and target domains through luminance alignment thus solves the problem of difficulty in obtaining accurate nighttime target annotations under low-light conditions; Second, a hybrid CNN-Transformer model is used as a detector. CSwin Transformer is used as a backbone network to extract multi-level image features and input these features into a feature pyramid network to facilitate multi-scale object detection. Finally, the Outlook attention module is introduced to address the loss of image details caused by insufficient illumination, enhancing the model's robustness under varying lighting conditions, shadows, and other complex environments. The results demonstrate that the proposed method achieved the mean Average Precision (mAP) of 50.0% on the BDD100K dataset, improving by 4.2 percentage points compared to the 2PCNet method. The mAP@0.5 reached 45.4% on the SODA10M dataset, improving by 0.9 percentage points compared to the SFA method.

Key words: unsupervised learning, object detection, nighttime domain adaptation, photometric alignment, CNN-Transformer model

摘要： 夜间目标检测相较于白天环境存在诸多挑战，受限于低光照条件和高质量标注数据的匮乏，难以提取目标特征，从而影响了夜间目标检测精度。为此，提出一种针对夜间图像的域适应目标检测方法。首先，设计一种夜间域适应光度对齐模块，将有标记的白天源域图像转换为对应的夜间目标域图像，即通过光度对齐弥合源域与目标域之间的差距，从而解决低光照条件下难以获取准确夜间目标注释的问题；其次，采用CNN-Transformer混合模型作为检测器，即以CSwin Transformer作为主干网络提取多层次的图像特征，并将提取特征输入特征金字塔网络中，以促进多尺度目标检测；最后，引入Outlook注意力解决由于光照不足导致的图像细节不明显问题，提升了模型在光照变化、阴影等复杂环境下的鲁棒性。实验结果表明，在公共数据集BDD100K上，所提方法平均精度均值(mAP@0.5)达到了50.0%，相较于2PCNet方法，mAP@0.5提高了4.2个百分点；在SODA10M数据集上，所提方法的mAP@0.5也达到了45.4%的精度，与SFA方法相比提高了0.9个百分点。

关键词: 无监督学习, 目标检测, 夜间域适应, 光度对齐, CNN-Transformer模型