An adaptive learning-based weakly supervised object detection via context awareness
Xiaoran Zeng, Zhenhua Li, Weidong Zhang · 2021
Weakly supervised object detection (WSOD) methods have become a powerful tool without fully labeled bounding boxes. In the field of object detection, however, there is still a certain gap in performance between the existing WSOD technique and the fully supervised object detection method. There exist two problems that are not common in fully supervised object detection including object ambiguity and falling into the local optima. To solve the object ambiguity problem, multiple instance learning with the self-training method is proposed in this paper, from which pseudo labels will be generated and gradually replace the training labels to locate objects more accurately during the training process. For the problem of falling into the local optima, a context awareness block is added into our module, which makes our network pay more attention to the background and context of our region of interest (ROI) object. Experiment results based on the dataset PASCAL VOC2007 and VOC2012 are carried out to demonstrate the effectiveness of the proposed approach.