Rethinking prediction alignment in one-stage object detection

doi:10.1016/j.neucom.2022.09.132

CASIA OpenIR > 中国科学院工业视觉智能装备工程实验室 > 精密感知与控制

	Rethinking prediction alignment in one-stage object detection
	Xiao, Junrui1,2 ; Jiang, He1,2 ; Li, Zhikai1,2 ; Gu, Qingyi1
发表期刊	NEUROCOMPUTING
ISSN	0925-2312
	2022-12-01
卷号	514 页码:58-69
通讯作者	Gu, Qingyi(qingyi.gu@ia.ac.cn)
摘要	Owing to their excellent performance and efficiency, one-stage detectors have been widely used in mul-timedia tasks, such as temporal action detection, object tracking, and video detection. However, misalign-ment between classification and regression branches limits the accuracy of the detector. Most existing works add an auxiliary branch or adopt a specific sample assignment strategy to alleviate this problem, but with little effect. In this paper, we attribute this to incomplete branch interactions and propose a comprehensive Predictive Aligned Object Detector (PAOD), which can better correlate two subtasks. Specifically, our proposed PAOD achieves a better trade-off between prediction-interactive and prediction-specific by adopting an Iterative Aggregation Module (IAM) and a Mutual Constraint Module (MCM). We also design an aligned label assignment with an adaptive metric and re-weighting mechanism to further narrow the misalignment between prediction heads. With negligible additional overhead, PAOD achieves 50.4 AP at single-model single-scale testing on the MS-COCO branch, which demonstrates the effectiveness of our proposal. Notably, PAOD consistently outperforms previous sota such as ATSS (47.7 AP), BorderDet (48.0 AP) and GFL (48.2 AP) by a large margin on COCO test-dev data -set, and achieves better performance than various dense detectors on Pascal VOC and CrowdHuman data -sets. Code is available at https://github.com/JunruiXiao/PAOD.(c) 2022 Elsevier B.V. All rights reserved.
关键词	Deep learning Object detection Detection head Label assignment
DOI	10.1016/j.neucom.2022.09.132
收录类别	SCI
语种	英语
资助项目	Scientific Instrument Developing Project of the Chinese Academy of Sciences ; [YJKYYQ20200045]
项目资助者	Scientific Instrument Developing Project of the Chinese Academy of Sciences
WOS研究方向	Computer Science
WOS类目	Computer Science, Artificial Intelligence
WOS记录号	WOS:000869405900004
出版者	ELSEVIER
引用统计	被引频次：8[WOS] [WOS记录] [WOS相关记录]
文献类型	期刊论文
条目标识符	http://ir.ia.ac.cn/handle/173211/50261
专题	中国科学院工业视觉智能装备工程实验室_精密感知与控制
通讯作者	Gu, Qingyi
作者单位	1.Chinese Acad Sci, Inst Automat, East Zhongguancun Rd, Beijing, Peoples R China 2.Univ Chinese Acad Sci, Sch Artificial Intelligence, Jingjia Rd, Beijing, Peoples R China
第一作者单位	中国科学院自动化研究所
通讯作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	Xiao, Junrui,Jiang, He,Li, Zhikai,et al. Rethinking prediction alignment in one-stage object detection[J]. NEUROCOMPUTING,2022,514:58-69.
APA	Xiao, Junrui,Jiang, He,Li, Zhikai,&Gu, Qingyi.(2022).Rethinking prediction alignment in one-stage object detection.NEUROCOMPUTING,514,58-69.
MLA	Xiao, Junrui,et al."Rethinking prediction alignment in one-stage object detection".NEUROCOMPUTING 514(2022):58-69.