Uncertain region mining semi-supervised object detection

Semi-supervised learning uses a small amount of labeled data to guide the model and a large amount of unlabeled data to improve its performance. Most semi-supervised object detection methods build a teacher-student architecture and train the student network with pseudo-labels generated by the teacher. To guarantee the accuracy of pseudo-labels, a high threshold value is always applied to filter out all of the low-scoring labels according to the inference results of the teacher. We argue that these discarded labels with low scores have more information for the current model and should be taken into consideration in the training phase. Further, we propose an uncertain region mining(URM) framework that utilizes these uncertainty low confidence labels. Especially, URM exploits the uncertain labels from two aspects: (1)Recalling the underlying correct labels: URM designs a fusion function that rectifies the outputs of the teacher with the student and recalls the high-quality pseudo-labels. (2)Avoiding the error information of the error labels: URM proposes a negative loss function that utilizes the uncertain labels without introducing error information. For the regression task, a new branch is attached to the detector to predict the localization scores of bounding boxes. Based on the predicted scores, we propose a re-weighting strategy that alleviates the noisy problem from the imprecise localization of the bounding boxes. Experiments on PASCAL-VOC, MS-COCO and DOTA datasets demonstrate the effectiveness of our proposed method.