A Mean-Teacher Framework for Source-Free Domain Adaptive Foggy Object Detection

Source-free domain adaptive object detection (SFOD) has achieved significant progress, Most SFOD methods leverage a Mean-Teacher (MT) self-training paradigm relying heavily on high-confidence Reliable Pseudo Labels (RPL). However, in foggy scenes, the loss of image details and blurring of object-background boundaries cause RPL to overlook small instances. Additionally, RPL ignore instances with low confidence due to the scarcity of training samples, resulting in biased adaptation toward familiar instances from the source domain. To address this limitation, we propose a MT framework for SFOD in foggy. At the feature extraction level, we propose a Weighted Bidirectional Feature Pyramid Network (WBiFPN), which focuses on key features and thereby alleviates the problems of image detail blurring and feature degradation caused by fog. To further mine positive samples, at the pseudo-label optimization level, we design a Pseudo-Label Graded Adaptation (PLGA) module, which adopts self-supervised and contrastive learning strategies for reliable and unreliable pseudo-labels respectively, enhancing the model's ability to mine potential positive samples and further improving the recognition and localization accuracy of small objects. The experimental results show that the average accuracy mean mAP reaches 38.7% in Cityscapes-Foggy Cityscapes scenarios, which is able to effectively cope with the task of target detection in foggy weather scenarios.