CAREER: Large-Scale Recognition Using Shared Structures, Flexible Learning, and Efficient Search

本研究探讨了适用于大规模视觉识别的共享表征、灵活的学习技术和高效的多类别推理方法。其目标是生产能够以不同精度准确描述广泛对象的视觉系统，而不是局限于识别几个预定义类别中的对象。主要的方法是设计对象表示，使新对象能够被现有对象理解，这使得学习可以用更少的例子和更快和更健壮的识别。研究包括三个主要部分：(1)为跨基本类别共享的对象设计外观和空间模型；(2)研究从详细和松散的注释和人类反馈的混合中学习的算法；(3)设计利用共享表示的高效搜索算法。该研究提供了更详细、更灵活、更准确的识别算法，适用于车辆安全、安保、助盲、家用机器人、多媒体搜索和组织等高影响应用。例如，如果车辆在路上遇到一头牛，视觉系统将定位这头牛及其头和腿，并报告“四条腿的动物，走在左边”，即使它在训练期间没有看到牛。这项研究还提供了一个独特的机会，让本科生参与研究，促进跨学科学习和合作，并参与外展。研究想法和结果通过科学出版物、发布的代码和数据集、公开演讲和高中生示范来传播。

This research investigates shared representations, flexible learning techniques, and efficient multi-category inference methods that are suitable for large-scale visual recognition. The goal is to produce visual systems that can accurately describe a wide range of objects with varying precision, rather than being limited to identifying objects within a few pre-defined categories. The main approach is to design object representations that enable new objects to be understood in terms of existing ones, which enables learning with fewer examples and faster and more robust recognition.The research has three main components: (1) Designing appearance and spatial models for objects that are shared across basic categories; (2) Investigating algorithms to learn from a mixture of detailed and loose annotations and from human feedback; and (3) Designing efficient search algorithms that take advantage of shared representations. The research provides more detailed, flexible, and accurate recognition algorithms that are suitable for high-impact applications, such as vehicle safety, security, assistance to the blind, household robotics, and multimedia search and organization. For example, if a vehicle encounters a cow in the road, the vision system would localize the cow and its head and legs and report ``four-legged animal, walking left'', even if it has not seen cows during training. The research also provides a unique opportunity to involve undergraduates in research, promote interdisciplinary learning and collaboration, and engage in outreach. Research ideas and results are disseminated through scientific publications, released code and datasets, public talks, and demonstrations for high school students.