IDO: Instance dual-optimization for weakly supervised object detection

doi:10.1007/s10489-023-04956

CORC > 自动化研究所 > 中国科学院自动化研究所 > 多模态人工智能系统全国重点实验室

	IDO: Instance dual-optimization for weakly supervised object detection
	Ren, Zhida 1,2; Tang, Yongqiang 2; Zhang, Wensheng 1,2
刊名	APPLIED INTELLIGENCE
	2023-08-29
页码	18
关键词	Deep learning Weakly supervised learning Object detection Multiple instance learning
ISSN号	0924-669X
DOI	10.1007/s10489-023-04956
通讯作者	Tang, Yongqiang(yongqiang.tang@ia.ac.cn)
英文摘要	Weakly supervised object detection (WSOD) has attracted significant attention in recent years, as it utilizes only image-level annotations to train object detectors and greatly reduces the labor and capital cost of fine labeling. Nevertheless, the absence of instance-level annotations leads to two phenomena: partial regions and missing instances. We believe these are mainly caused by two issues: 1) Noisy instances exist in the training samples, which can confuse the detector. 2) Global salient information is missing, resulting in little attention being received in the low-confidence region. To solve the above two problems, we propose an instance dual-optimization framework called IDO. First, an instance-wise selection strategy (IWSS) based on curriculum learning is proposed for instance denoising and for improving the robustness of the model. Second, CAM-generated spatial attention (CGSA) is carefully designed to optimize the features of instances. Without introducing additional hyperparameters, our CGSA complements the low class-confidence region with more global salient information, which assists the model in acquiring a more complete region of the target and identifying more neglected targets. Finally, we empirically demonstrate that our proposal can achieve comparable results to those of other state-of-the-art methods on PASCAL VOC 2007, PASCAL VOC 2012, and MS COCO.
资助项目	National Key Research and Development Program of China[2020AAA0109600] ; National Natural Science Foundation of China[62106266] ; National Natural Science Foundation of China[62173328] ; National Natural Science Foundation of China[62006139] ; National Natural Science Foundation of China[61976213]
WOS研究方向	Computer Science
语种	英语
出版者	SPRINGER
WOS记录号	WOS:001060433700001
资助机构	National Key Research and Development Program of China ; National Natural Science Foundation of China
内容类型	期刊论文
源URL	[http://ir.ia.ac.cn/handle/173211/54155]
专题	多模态人工智能系统全国重点实验室
通讯作者	Tang, Yongqiang
作者单位	1.Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing, Peoples R China 2.Chinese Acad Sci, Inst Automat, State Key Lab Multimodel Artificial Intelligence S, Beijing, Peoples R China
推荐引用方式 GB/T 7714	Ren, Zhida,Tang, Yongqiang,Zhang, Wensheng. IDO: Instance dual-optimization for weakly supervised object detection[J]. APPLIED INTELLIGENCE,2023:18.
APA	Ren, Zhida,Tang, Yongqiang,&Zhang, Wensheng.(2023).IDO: Instance dual-optimization for weakly supervised object detection.APPLIED INTELLIGENCE,18.
MLA	Ren, Zhida,et al."IDO: Instance dual-optimization for weakly supervised object detection".APPLIED INTELLIGENCE (2023):18.