中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
首页 机构 成果 学者
  
  1. CAS IR Grid
  2. 自动化研究所
  3. 中国科学院自动化研究所
  4. 多模态人工智能系统全国重点实验室
  5. 机器人应用与理论组
Active domain adaptation for semantic segmentation via dynamically balancing domainness and uncertainty

文献类型：期刊论文

作者Zhang, Siqi1,2; Zhang, Lu1; Liu, Zhiyong1,2,3
刊名IMAGE AND VISION COMPUTING
出版日期2024-08-01
卷号148页码:11
关键词Active learning Domain adaptation Semantic segmentation Transfer learning
ISSN号0262-8856
DOI10.1016/j.imavis.2024.105132
通讯作者Liu, Zhiyong(zhiyong.liu@ia.ac.cn)
英文摘要Active domain adaptation aims to enhance model adaptation performance by annotating a limited number of informative unlabeled target data. Traditional active learning strategies for semantic segmentation often neglect the presence of domain shifts, resulting in suboptimal results in domain adaptation scenarios. In this paper, we present a novel active domain adaptation approach for semantic segmentation that maximizes segmentation performance under domain shifts with a limited number of queried target labels. To recognize the most valuable samples for labeling, we introduce a new acquisition strategy. This strategy leverages a target domainness map to identify the most informative samples for reducing the domain gap and employs region-aware prediction uncertainty to explore ambiguous samples. Meanwhile, to optimize the efficiency of the acquisition strategy, we dynamically adjust the balance between prediction uncertainty and target domainness over the selection rounds. To further bolster adaptation performance, a smooth loss function is employed for the target data, which promotes consistency in local predictions. Extensive experiments on two benchmarks, GTAV -* Cityscapes and SYNTHIA -* Cityscapes, demonstrate that our method surpasses existing active domain adaptation methods for semantic segmentation. Moreover, it achieves comparable results to supervised performance with only 5% annotations in the target domain, validating the effectiveness of our method.
资助项目National Key Research and Development Plan of China[2020AAA0108902] ; NSFC[62206288]
WOS研究方向Computer Science ; Engineering ; Optics
语种英语
WOS记录号WOS:001259799200001
出版者ELSEVIER
资助机构National Key Research and Development Plan of China ; NSFC
源URL[http://ir.ia.ac.cn/handle/173211/59170]  
专题自动化研究所_复杂系统管理与控制国家重点实验室_机器人应用与理论组
通讯作者Liu, Zhiyong
作者单位1.Chinese Acad Sci, Inst Automat, State Key Lab Multimodal Artificial Intelligence S, Beijing, Peoples R China
2.Univ Chinese Acad Sci, Sch Artificial Intelligence, Beijing, Peoples R China
3.Nanjing Artificial Intelligence Res IA, Nanjing, Jiangsu, Peoples R China
推荐引用方式
GB/T 7714		 Zhang, Siqi,Zhang, Lu,Liu, Zhiyong. Active domain adaptation for semantic segmentation via dynamically balancing domainness and uncertainty[J]. IMAGE AND VISION COMPUTING,2024,148:11.
APA Zhang, Siqi,Zhang, Lu,&Liu, Zhiyong.(2024).Active domain adaptation for semantic segmentation via dynamically balancing domainness and uncertainty.IMAGE AND VISION COMPUTING,148,11.
MLA Zhang, Siqi,et al."Active domain adaptation for semantic segmentation via dynamically balancing domainness and uncertainty".IMAGE AND VISION COMPUTING 148(2024):11.

入库方式： OAI收割

来源：自动化研究所

浏览0
下载0
收藏0
其他版本

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。