Object recognition is one of the most important problems in computer vision. While researchers have worked on this problem for over thirty years, vision systems are still unable to recognize many common objects in cluttered images. The PI proposes to address this problem by developing new hierarchical models and efficient search algorithms for recognition.
Hierarchical models represent objects using parts which are themselves defined in terms of subparts. Moreover, the subparts may be recursively defined in terms of smaller components. This hierarchical organization can efficiently encode important relationships among the components that make up an object. Another important property of hierarchical models is that components can be shared among different object models. This is useful for being able to quickly recognize which of many possible objects are present in an image. It is also important for learning models from small datasets. Finally, in the most general types of models the structure of an object may be specified by a grammar instead of being fixed in advance. The number of parts that make up an object may be variable and there may be choice among different parts that can go in a particular place. All of these aspects make hierarchical models incredibly expressive.
Algorithms for object recognition typically search over large spaces encoding the pose of an object, or over correspondences between model features and features extracted from an image. The PI will develop efficient optimization algorithms for solving these problems. This will be accomplished by exploiting the structure of the search spaces defined by general classes of hierarchical models.
Broader significance and importance: Object recognition has many important practical applications, including in robotics, human-computer interaction, image retrieval, security systems and medical image analysis. Research in object recognition can also play an important role in our understanding of human perception and intelligence. The proposed research will draw upon ideas from diverse areas such as computer vision, theoretical computer science, natural language understanding and mathematics.
URL: http://people.cs.uchicago.edu/~pff/hierarchical
The project has involved the development of algorithms for object recognition in images and video. There were two main directions of research. The first involves deformable part-based models for object detection. The second involves stochastic grammars for shape modeling, recognition and detection. Our work on object detection has led to a state-of-the-art object detection system. The system has had a significant impact in the field and is widely used by other researchers in academia and industry. We received the PASCAL 'Lifetime Achievement' prize in 2010 in recognition of our contribution, both due to its innovation and success and also its free distribution. We have also developed a more general framework for object detection based on the notion of stochastic grammars. Object detection grammars define a general class of part-based models that allow for efficient inference and discriminative training. An important outcome of the project is a free software package for object detection using state-of-the-art algorithms. The software package is available from the PI's website. This system has become a significant resource for the computer vision community; both for research and education. A significant amount of the current research in object recognition builds on this system. Students taking computer vision courses in many universities have used the source code we distribute in their own projects. The project provided funding for three PhD students that worked on several aspects of the research. The research was presented at several academic workshops and conferences. The research has led to several publications that are highly cited. URL: http://cs.brown.edu/~pff/latent
