“…Predictions are then filtered to iteratively generate strong pseudo-labels, enabling unbiased updates to the student network. Furthermore, to reduce the image-level discrepancy between the teacher and student models, enhance domain consistency, and enrich the diversity of training inputs, we adopt Contrastive Unpaired Translation (CUT) [10], a state-of-the-art unsupervised image-to-image translation method. CUT operates by transforming source domain images into the stylistic characteristics of the target domain, thereby generating pseudo-images in an offline manner.…”