Using Geometric Information in recognition and scene analysis
Geometry, defined broadly to include any partial representation of shape, is critical for relating models of objects and scenes with perceived images. However, it is still difficult to use geometric information effectively. I will discuss several aspects of using geometric information for recognition and scene analysis at different levels, from weak shape features to more complete 3D data. Weak geometric features include contour fragments describing objects' shapes and relative locations of local point features. The difficulty in using this type of information is that it is very ambiguous. Techniques based on targeted use of quantization and on matching approaches can be used to deal with such partial data. Intermediate geometric representation include scene geometric context in which major surfaces are identified and labeled, thus constraining the interpretation of the rest of the image. Recent progress in the use of geometric context shows how to effectively use these constraints. Finally, full 3D data, when available, can be use to reason directly about object shapes and about local 3D distribution of data.delivery. We will end the talk with current and future research directions.
Martial Hebert in a professor in the Robotics Institute, Carnegie Mellon University. His interests are in computer vision—especially object recognition, scene understanding, and video analysis— processing of 3-D data for building 3-D models, scene analysis, and perception for autonomous mobile systems. His current projects include the development of techniques for recognizing categories and objects in images, for detecting common events in video sequences, and for building 3-D representations of dynamic environments for unmanned mobile systems. In the area of object detection and localization for manipulation, recent work includes 2D/3D recognition techniques used in the NSF ERC "quality of Life Technology" . In the area of mobile robotics, recent projects include the development of detection, tracking, and prediction techniques for unmanned vehicles as part of the Robotics Collaborative Technology Alliance.