Cropping-classification-based semi-supervised learning for object detection

In current semi-supervised object detection techniques, one-stage detectors typically realize only limited improvements when compared to two-stage detectors. The limitation is largely due to the assignment of incorrect labels to samples during the label allocation phase, as the process is influenced by inaccurate pseudo-labels. To tackle the problem, we propose a Cropping-Classification Semi-supervised Learning (CCSL) for one-stage detectors. After detection, we employ a Cropping-Classification module (CCM) designed to refine inaccurate pseudo-labels and to bolster the precision of localization. Extensive experiments on all SSOD settings of both COCO and Pascal VOC benchmark datasets show that our CCSL method outperforms all state-of-the-art one-stage methods. The experiments on the DIOR dataset show that CCSL has good generalization ability.