A Joint Framework based on PU Learning and Active Learning for Object Detection
Yue Li, Minzheng Jia · 2023
The problem of object detection is a critical issue in the field of image processing, and the quality of labels can significantly affect the performance of detectors. In recent years, many deep learning frameworks have achieved good detection results. However, in many practical scenarios of object detection, collecting complete annotation is a time-consuming and costly process. Therefore, how to use partially labeled samples to solve the object detection problem has attracted the attention of many researchers. We proposed a new joint training framework based on PU (Positive-Unlabeled) learning and active learning to solve the above problem, named the PU_AL learning method. This framework is based on the Faster RCNN two-stage detector, and we introduce the PU learning method to solve the problem of incomplete annotations, adapting the traditional semi-supervised learning problem to a PU learning problem. We also introduce the loss branch prediction method based on Active Learning to improve the selective labeling method of sample selection. Experiment results demonstrate that the proposed framework outperforms the baseline methods in varying degrees of missing annotation.