CMU FRC talk: Recognizing Things in Images

Speaker: Martial Hebert, Professor, Robotics, Carnegie Mellon University
Date: Thursday, March 16, 2006

Abstract: Finding things (objects, regions, events) in images or video is the main objective of computer vision. A common view of this problem is to start with tentative labeling of parts of the image as possible locations of objects or possible types of regions, followed by reasoning about context and relations between these elements. We have been working on tools for recognition that are very effective for this type of tasks. In particular, we have developed tools for representing relations beween scene elements (features, objects, regions..) and for enforcing geometric constraints between elements.

I'll review some of the recent results in using relations and context between image elements for recognition and classification. The applications include finding salient structures in images, recognizing individual objects, and segmenting the image into labeled regions. Most of the examples deal with single images but the techniques can be used also for more recognizing things in video sequences and I will show some results in this area from a recently completed project. Applications include detection of useful landmarks, object localization for navigation and manipulation, and surveillance.

Speaker Bio: Martial Hebert is Professor at the Robotics Institute, Carnegie Mellon University. He has led many major computer vision and robotics projects, funded by DARPA, NASA, NSF, ONR, DOE, and industry. Prof. Hebert has worked in multiple areas of robotics: computer vision, autonomous mobile robots, and sensors. His current research interests include object recognition in images, video, and range data, scene understanding using context representations, and model construction from images and 3-D data. His group has explored applications in the areas of autonomous mobile robots, both in indoor and in unstructured, outdoor environments, automatic model building for 3D content generation, and video monitoring. He has published more than 150 technical papers and reports in these areas.