Co‐teaching based pseudo label refinery for cross‐domain object detection

Object detection is one of the main tasks in computer vision and has made great progress recent years. However, performance target detectors significantly dropped by differences between existing datasets application scenarios, leading to so-called domain shift problem. To address such an issue, a novel co-teaching based pseudo label refinery framework for cross-domain object developed, which cooperates with two models select data from each other. This strategy can effectively purify predicted labels resist noisy labels. Specifically, consists encoders (i.e. structure encoder global encoder), classifiers discriminator, used extract structural features that are not disturbed colour, complete discriminant features. The followed classifier. In training, labelled source samples first trained, so it initial recognition ability. Then assigned classifier following fine-tune pre-trained on obtain refined data. With labels, further optimised domain. During this process, proposal cross use promote mutual transfer complementary capabilities encoders. Moreover, residual channel attention block (RCA) embedded salient designed pay more regions. Extensive experiments demonstrate developed generate clean unlabelled boost detection. code available at http://www.msp-lab.cn:1436/msp/cbplr-master.