Incremental Few-Shot Object Detection with scale- and centerness-aware weight generation

doi:10.1016/j.cviu.2023.103774

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

	Incremental Few-Shot Object Detection with scale- and centerness-aware weight generation
	Zhang, Lu 4,5; Yang, Xu 5; Qi, Lu 2; Zeng, Shaofeng 5; Liu, Zhiyong 1,3,4,5
刊名	COMPUTER VISION AND IMAGE UNDERSTANDING
	2023-10-01
卷号	235 页码:9
关键词	Incremental learning Few-shot learning Object detection
ISSN号	1077-3142
DOI	10.1016/j.cviu.2023.103774
通讯作者	Liu, Zhiyong(zhiyong.liu@ia.ac.cn)
英文摘要	Our work targets the task of Incremental Few-Shot Detection (iFSD), which requires quickly detecting objects of novel classes after seeing only a few training examples, while maintaining the ability to detect old classes. Different from re-training a new detector to incorporate few-shot novel categories, we tackle iFSD in a direct weight generation way. Specifically, we propose to reorganize base weights by leveraging responses of novel region features, hence transferring learned information from base weights to the novel but relevant ones. To obtain representative region features for the novel weight generation, two strategies are further presented. The scale-aware strategy respects the scale variation in object detection, and assigns each object to its corresponding pyramid to obtain the scale-specific region feature. The centerness-aware strategy reorganizes region features in the spatial range, since the localization quality of each pixel on the feature map varies. Extensive experiments are conducted on the challenging MS COCO and PASCAL VOC datasets, and our method achieves promising performances under various settings with a direct weight generation fashion. Furthermore, our generated weights can be used as good initialization for fine-tuning to induce faster convergence, thus mitigating the performance drop on old classes and improving overall results in the incremental setting.
资助项目	Na-tional Key Research and Development Plan of China[2020AAA0108902] ; NSFC[62206288] ; Youth Innovation Promotion Association CAS
WOS研究方向	Computer Science ; Engineering
语种	英语
出版者	ACADEMIC PRESS INC ELSEVIER SCIENCE
WOS记录号	WOS:001047035100001
资助机构	Na-tional Key Research and Development Plan of China ; NSFC ; Youth Innovation Promotion Association CAS
内容类型	期刊论文
源URL	[http://ir.ia.ac.cn/handle/173211/53993]
专题	多模态人工智能系统全国重点实验室
通讯作者	Liu, Zhiyong
作者单位	1.Zhongguancun East Rd 95, Beijing 100190, Peoples R China 2.Chinese Univ Hong Kong, Hong Kong, Peoples R China 3.Nanjing Artificial Intelligence Res IA, Nanjing 211100, Jiangsu, Peoples R China 4.Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing 100049, Peoples R China 5.Chinese Acad Sci, Inst Automat, State Key Lab Multimodal Artificial Intelligence S, Beijing 100190, Peoples R China
推荐引用方式 GB/T 7714	Zhang, Lu,Yang, Xu,Qi, Lu,et al. Incremental Few-Shot Object Detection with scale- and centerness-aware weight generation[J]. COMPUTER VISION AND IMAGE UNDERSTANDING,2023,235:9.
APA	Zhang, Lu,Yang, Xu,Qi, Lu,Zeng, Shaofeng,&Liu, Zhiyong.(2023).Incremental Few-Shot Object Detection with scale- and centerness-aware weight generation.COMPUTER VISION AND IMAGE UNDERSTANDING,235,9.
MLA	Zhang, Lu,et al."Incremental Few-Shot Object Detection with scale- and centerness-aware weight generation".COMPUTER VISION AND IMAGE UNDERSTANDING 235(2023):9.