Mean teacher based on class prototype contrast for domain adaptive object detection

Unsupervised domain adaptive object detection (UDAOD) aims to effectively apply the detector trained on a labeled (source domain) and an unlabeled (target domain) dataset to the target domain. The mean teacher framework has demonstrated good applicability and wide application in this task. However, influenced by the difference between the two domains, the teacher model often generates many false positive objects. The pseudo-labels cannot sufficiently include all classes of objects in an image because of single-threshold filtering, causing the model to perform poorly in detection tasks. Therefore, we propose a new student-teacher framework, the mean teacher, which is based on class prototype contrast (PCMT). Utilizing class prototypes to preserve the features that are common in objects of the same class to address the problem of significant feature differences that may exist between these objects. Then, the class prototypes are applied to contrastive learning, so that the model can distinguish various classes more accurately while align the features of the same class across domains. In addition, we design a pseudo-label filtering method based on bounding box localization to retain potentially valid pseudo-labels. Experiments show that PCMT achieves superior performance under different domain adaptive conditions. For the Cityscapes →  BDD100K dataset, we obtain the best mean average precision (mAP) of 43.5%, which is 5.0% greater than the state-of-the-art (SOTA).