CAREER: Image Variability Decomposition for Recognition, Reconstruction, and Tracking

职业：用于识别、重建和跟踪的图像变异分解

• 批准号: 0234606
• 负责人: Peter Belhumeur
• 金额: $ 9.22万
• 依托单位: Columbia University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2002
• 资助国家: 美国
• 起止时间: 2002-01-01 至 2004-06-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0234606&HistoricalAwards=false
• 关键词: CAREER; Image; Variability; Decomposition; Recognition

This research investigates how changes in illumination, pose, and shape of an object produce changes in the observed images. The central idea is that the variability in the images of an object can be decomposed into component parts - illumination, pose, and shape - each of which, when analyzed separately, is well behaved. This paradigm, termed ``Image Variability Decomposition,'' differs substantially from the appearance-based paradigm in that by decomposing the variability, one is able to uncover generative structures to the set of images over each source of the variability. Thus, unlike appearance-based methods, it is not necessary to have seen the object under all of the seemingly infinite possible permutations of lighting conditions, pose, and shape. Instead, each component of the image variability is explicitly modeled. This research is applied to problems in which this variability plays an important role: face recognition, change detection, structure from motion, and visual tracking. The education component of this award focuses on laboratory and project intensive teaching. Students, both graduate and undergraduate, are not only taught about known unsolved problems, but are encouraged to consider posing new problems in computer vision, robotics, and pattern recognition. Students are also rewarded for finding new applications in new domains. The education activities will center around the construction of a computer vision laboratory within Yale University's newly formed Center for Computational Vision and Control, the design of laboratory intensive courses on both computer vision and pattern and object recognition, and the completion of a textbook on computer vision.

这项研究调查了物体的照明、姿势和形状的变化如何在观察到的图像中产生变化。其核心思想是，物体图像中的可变性可以分解为组成部分--照明、姿势和形状--当单独分析时，每一部分都表现良好。这一范式称为“图像可变性分解”，它与基于外观的范式有很大的不同，因为通过分解可变性，人们能够揭示可变性的每个来源上的图像集的生成结构。因此，与基于外观的方法不同，没有必要在光照条件、姿势和形状的所有看似无限可能的排列下看到对象。取而代之的是，对图像可变性的每个分量进行显式建模。这项研究被应用于这种变异性发挥重要作用的问题：人脸识别、变化检测、运动结构和视觉跟踪。该奖项的教育部分侧重于实验室和项目密集教学。学生，包括研究生和本科生，不仅被教授已知的未解决的问题，而且被鼓励在计算机视觉、机器人和模式识别方面提出新的问题。学生还会因为在新领域发现新的应用程序而获得奖励。教育活动将围绕在耶鲁大学新成立的计算视觉与控制中心内建立一个计算机视觉实验室，设计计算机视觉、模式和物体识别的实验室强化课程，以及完成一本计算机视觉教科书。