CAREER: Weakly Supervised Recognition

能够在大型图像和视频存储库中检测和识别物体的系统设计将在生命科学、监视和执法、娱乐、广告和版权保护等领域取得重大进展。虽然在过去的几十年里，这一领域取得了重大进展，但这种系统的设计仍然需要大量的专家知识和体力劳动。该项目为包含识别模块库的识别系统的长期愿景奠定了基础，这些识别模块库可由幼稚用户完全训练，在人工数据预处理和计算复杂性方面要求最低。从技术角度来看，该项目解决了实现这一目标道路上的两个基本障碍：1)当前分类器对精心组装和预处理的训练集的依赖，以及2)最先进分类体系结构的训练复杂性。第一个问题是通过引入一个新的统计学习框架来解决的，该框架由训练集的弱监督组装表示，它结合了判别视觉显著性和图像匹配的元素，以自动组装检测和识别所需的训练集的过程。第二个问题是通过引入新的、计算效率高的、用于设计大边界分类器级联的增强方法来解决的，该方法支持模板和基于星座的对象表示。在教育方面，该项目将通过开设新课程和开发一个软件库，作为视觉资料检索课程的教学辅助工具，促进提高识别问题的覆盖面。该图书馆的设计还将为来自代表性不足的背景的学生提供研究机会，其规模远远超出通常在本科阶段可以找到的范围。

The design of systems that can detect and recognize objects in large image and video repositories will enable significant developments in areas as diverse as life sciences, surveillance and law enforcement, entertainment, advertisement, and copyright protection, among others. While significant progress has been achieved in this area over the last decades, the design of such systems still requires vast amounts of expert knowledge and manual labor. This project lays the foundation for a long-term vision of recognition systems containing banks of recognition modules fully trainable by naive users, with minimal requirements in terms of manual data pre-processing and computational complexity. From a technical standpoint, the project addresses two fundamental barriers in the path to this objective: 1) the dependence of current classifiers on carefully assembled and pre-processed training sets, and 2) the training complexity of state-of-the-art classification architectures. The first is addressed through the introduction of a new statistical learning framework, denoted by weakly supervised assembly of training sets, which combines elements of discriminant visual saliency and image matching to automate the process of assembling the training sets required for detection and recognition. The second is addressed through the introduction of new, and computationally efficient, boosting methods for the design of cascades of large-margin classifiers, with support for both template and constellation-based object representations. At the educational level, the project will contribute to the advancement of the coverage of the recognition problem, through the introduction of new courses, and the development of a software library to be used as a teaching aid in visual information retrieval courses. The design of this library will also provide research opportunities for students from underrepresented backgrounds at a scale well beyond what can usually be found at the undergraduate level.