标注受限的深度目标检测算法研究

近年来，基于海量标注数据的深度目标检测算法的研究，取得了瞩目的成果。但是深度目标检测算法经常处于数据标注受限和不足的窘境，出现性能滑坡。针对这些问题，本申请拟对数据标注受限的深度目标检测算法展开研究。主要研究内容包括：一、基于一致性和差异性建模，进行标注的挖掘和修正，对基于图像级标注的弱监督目标检测算法进行研究; 二、研究如何充分利用标注数据，捕捉标注和无标注数据交互信息，构建半监督目标检测新方法，建立半监督目标检测统一框架。三、基于自监督学习和实例级别匹配，提升特征质量，进行小样本目标检测算法的研究。四、研究如何在目标域无标注数据的场景下，基于领域自适应和领域泛化，对目标检测模型的知识进行跨域迁移。申请书详细论述了研究内容的意义，提出了切实可行的研究方案和技术路线。项目组具备较强的研究基础和完善的工作条件。预计本项目顺利开展会推动整个目标检测领域的理论和应用上的革新。

In recent years, the research of deep object detection based on massive labeled data has achieved remarkable success. However, deep object detection model is often in the dilemma of limited labeling data, resulting in performance degradation. ..To solve this problem, we would like to investigate the deep object detection algorithms with limited annotations. Firstly, we’d like to study the weak supervised object detection algorithm based on image-level labeling. We would mine and modify the pseudo annotations by using the consistency and difference modeling technique. Second, we’d like to study how to make full use of the labeled data, capturing the interactive information of labeled and unlabeled data, constructing new algorithms of semi-supervised object detection, and establishing a unified framework of semi-supervised object detection. Third, we’d like to improve the feature quality based on self-supervised learning and case-level matching to carry on the research of few shot detection algorithm. Fourth, we’d like to study how to transfer the knowledge of the detection model across domains to target domain without annotations based on domain adaptation/domain generalization. ..We discuss the motivations and values of these studies and propose feasible research schemes in detail. This project will be organized in an efficient way. The group members have great research backgrounds and have accumulated a lot of research experience in related fields. We expect to coin a lot of novel deep object detection algorithms with limited annotations, and conduct some applications.