Active Learning Method Based on Pseudo-labeling

As an important fundamental task in the field of computer vision, object detection has a wide range of applications. The performance of object detection not only depends on excellent network structures but also requires the support of large and rich annotated datasets. However, annotating data requires a considerable amount of personnel and funding, and the quality of the annotated data greatly affects the accuracy of object detection algorithms. In order to solve the problem of labeling cost, researchers, we proposed a multi-scale feature fusion method is designed to fuse and reduce the dimensionality of the multi-scale features of unlabeled samples, while retaining the main features of the samples and reducing computational burden. Clustering is used to apply pseudo-label classification to unlabeled samples, and information entropy is introduced to calculate the uncertainty of unlabeled samples, considering both uncertainty and diversity of the unlabeled samples. The experiment results on Pascal VOC data set and MS COCO data set prove that our method is equivalent to the latest methods. The ablation experiment proves that our method can effectively improve the labeling efficiency and can achieve the effect that would otherwise be achieved in a larger data set.