A Semi-supervised Object Detection Learning Method under Queue Smoothing Pseudo-label Supervising and Embedding Consistency Constraint

Abstract Semi-supervised object detection is an effective solution to balance the manual annotation cost and model performance in practical application. However, two major types of semi-supervised learning approaches based on the pseudo-labeling supervising and the consistency constrained self-supervising exist some limitation with the low confidence of pseudo-label and the suboptimal feature learning for specific-task respectively. To overcome the above limitation, we proposes an improved semi-supervised object detection learning method under queue smoothing pseudo-label supervising and embedding consistency constraint learning method. In detail, taking the teacher-student framework as the base model, two paralleled transforming modules i.e. a classification head and a embedding projection layer are constructed after the feature encoder. With the different data augmentation exerting on inputs at the teacher and the student module respectively, a pair of class-vector and embedding are obtained simultaneously for each proposal. Subsequently, under the smoothness assumption of class prediction probability within the same cluster, a class-vector is updating weighting with smooth constraints calculating in the embedding similarity between memorized neighboring samples and corresponding pseudo-label is corrected with the increasing confidence then. Furthermore, dual constraints are constructed based on pseudo-label supervising and consistency between the class-vector and embedding matrix together with a few label data for guiding the semi-supervised object detection learning. The experiments on the MS-COCO and PASCAL VOC datasets demonstrate that the proposed method outperforms the baseline method and several mainstream semi-supervised learning methods with the highest mAP.