Weakly Supervised Object Detection with Pyramid Squeeze Attention and Candidate Box Screening Algorithm
Wanchuan Jiang, Ling Ou · 2023
Weakly supervised object detection is mainly applicable when there is no instance-level category annotation. Due to the lack of annotations, it is difficult for weakly supervised methods to accurately predict the location of objects. Existing methods tend to converge to the most salient local regions of objects. To address this problem, we propose a weakly supervised object detection framework. In our framework, we add the Pyramid Squeeze Attention Module to the backbone feature extraction network to improve the localization ability of the network, thus enabling the network to consider the global features of objects. To further improve detection accuracy and obtain better candidate boxes from weakly supervised object detection networks as a pseudo-GT for regression and classification, we propose a Candidate Box Screening Algorithm (CBSA). Extensive experiments are run on the PASCAL VOC 2007 dataset under a single model. The method achieves a detection accuracy of 51.7% on this dataset. Improved detection accuracy compared to recent weakly supervised object detection methods.