Weakly supervised target detection based on spatial attention

• DOI: 10.1007/s44267-024-00037-y
• Publication Date: 2024-02-04T04:31:00Z
• Authors: Wenqing Zhao; Lijiao Xu
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology

Abstract Due to the lack of annotations in target bounding boxes, most methods for weakly supervised detection transform problem object into a classification candidate regions, making it easy detectors locate significant and highly discriminative local areas objects. We propose weak monitoring method that combines attention erasure mechanisms. The uses maps search with higher discrimination within then an mechanism erase region, forcing model enhance its learning features weaker discrimination. To improve positioning ability detector, we cascade network fully network, jointly train through multi-task learning. Based on validation trials, category mean average precision (mAP) correct localization (CorLoc) two datasets, i.e., VOC2007 VOC2012, are 55.2% 53.8%, respectively. In regard mAP CorLoc, this approach significantly outperforms previous approaches, which creates opportunities additional investigations identification algorithms.