基于稳健学习机制的复杂场景弱监督目标检测方法研究

Object detection is one of the most fundamental tasks in computer vision community. In order to relieve the time, energy and money that human cost for annotating the training data, weakly supervised object detection has become an urgent research direction in this field. Based on the fact that current weakly supervised object detection methods cannot handle complex scenes well, this project proposes a novel idea which builds weakly supervised object detection framework based on the robust learning mechanism. Such robust learning mechanism is able to overcome the challenging learning ambiguity issue inherent in visual learning under the weak supervision. Based on our previous works published on IEEE TPAMI and CVPR, this project focuses on four major parts: 1) Explore prior-knowledge that can be easily acquired and is highly-informative for the weakly supervised object detection task and develop effective and robust ways to involve it into the weakly supervised learning procedure; 2) Establish the cooperation scheme of related tasks under the weak supervision and develop the joint weakly supervised deep learning framework to simultaneously address the weakly supervised object localization and segmentation tasks; 3) Increase the robustness of the weakly supervised object detection approach to the noise data and its capacity to leverage the unlabeled data; 4) Construct a large-scale dataset containing complex visual scenes for algorithm testing and comparison. This project will be conducted in a manner of international cooperation. It aims for innovative achievements in both theory and application and will also be very attractive for other relevant domains of image understanding.

目标检测是计算机视觉领域中最根本的任务之一。为了减轻在标注训练数据中人们消耗的时间、精力和金钱，弱监督目标检测成为了该领域亟待研究的方向。针对现有方法在复杂场景下效果不佳的现状，本项目提出利用稳健学习理论克服弱监督学习下所固有的学习不确定性，并以此构建基于稳健学习机制的弱监督目标检测框架的新思路。结合申请人前期工作（成果发表于IEEE TPAMI, CVPR等），重点研究以下四个内容：1）针对弱监督目标检测任务探索容易获取且具有指导意义的先验知识，并将其有效、稳健地引入弱监督学习过程；2）建立弱监督下的多任务协作机制，构建深度联合学习框架来同时解决弱监督目标检测与分割问题；3）增加弱监督目标检测方法对噪声数据的鲁棒性以及对无标签数据的可用性；4）构建包含复杂场景的大规模数据库用于算法测试和比较。本项目采用国际合作研究方式，力争在算法理论和应用方面取得创新成果，推动图像理解等相关领域的发展。

针对现有方法在复杂场景下遭遇的标注瓶颈，本项目提出利用稳健学习理论克服弱监督学习下所固有的学习不确定性，并以此构建基于稳健学习机制的弱监督目标检测框架的新思路。取得的代表性工作包括：1）提出了自步-课程学习机制以实现弱监督学习与步进学习联合建模下的鲁棒弱监督学习机制；2）提出了基于区域搜索范式的弱监督目标定位框架；3）提出了基于空间位置上下文和相关语义上下文的多实例推理框架；4）建立了联合进行弱监督目标检测与分割任务的稳健性学习方法；5）提出了分割-三维重建交互引导的弱监督学习框架；6）构建了广义弱监督目标检测公开测试集；7）建立了首个面向遥感影像分析的零样本目标检测数据集。 ..本项目研究产生了一批高水平研究成果，发表高水平学术论文27篇，其中领域重要的国际期刊论文22篇，包括IEEE TPAMI在内的IEEE汇刊系列14篇，本领域重要国际会议论文5篇，中文期刊论文1篇；申请/授权国家发明专利7项；培养博士生5名、硕士生4名、博士后1名；项目负责人获得2021年IEEE Transactions on Circuits and Systems for Video Technology最佳论文奖（IEEE TCSVT Best Paper Award），入选中国博士后创新人才支持计划、吴文俊人工智能优秀青年奖、科睿唯安“全球高被引科学家”，获ACM China SIGAI优博奖。