Object category learning and recognition

点击翻译按钮获取中文摘要

Computer vision systems are starting to be used for many important applications, such as warning drivers of impending accidents, diagnosing disease from medical images, or allowing robots to perform useful tasks in household environments.  However, the capabilities of human vision remain far beyond what computer vision systems can do.  We are able to look at a scene, photograph, or drawing and immediately recognize the objects that are present.  Computer vision has made dramatic advances within the past few years, but has only reached the point at which small well-defined object catgories (such as "faces" or "cars") or larger libraries of specific objects (such as a database of particular buildings) can be reliably recognized.  The research in this proposal aims to expand the capabilities of computer vision systems so that they can generalize to much broader categories of objects and scenes, and thereby open up a broad range of new applications.  The best current approaches to object class recognition are based on using many local features corresponding to scale-invariant patches of an image.  We propose to extend the useful set of features to a much broader class of image descriptions.  We will develop features based on local contours, shapes, and textures.  A particular area of investigation will be feature classes that are robust to background clutter because they respond only to features that are likely to be part of the object while ignoring regions outside its bounding contours.  We intend to examine a range of learning methods, based upon AdaBoost, SVM, or sparse logistic regression, that can provide feedback on selecting appropriate image features for particular object categories.  We will also work on new methods for pose clustering and verification.  Our goal is to combine all of these new components into a complete system for object class recognition, opening up important new applications for computer vision.