Carnegie Mellon Robotics Institute
I have always been interested in understanding how to construct systems capable of highly skilled perception. The problem combines engineering with proof of validity and generality. Much of my work involves inventing and cataloguing devices of interest to a theory of skilled perception. A great wealth of well defined material exists in commercial and manufacturing applications. These applications typically allow a definition of success, completeness, and autonomy. Does the device serve a useful and adaptive role? Is it a complete device with no assumptions, and is it perfectly autonomous?
Automatic visual inspection: Many, if not most, manufacturing activities lead to products whose appearance is critical to the rated quality of product. A particularly interesting and economically important case is the control of print quality on offset and Intaglio printing presses. Here are cases of a manufactured product which depends almost completely on manufactured appearance, and there are single machines which can be servoed to alter the physical appearance of the manufactured goods.
Automatic document interpretation: Inspecting a document for quality is related to the automatic interpretation of documents from camera images. We have studied this problem in a number of documentation areas: Scanning film and converting the film to N/C control information for photoplotters, automatically scanning and interpreting oil well logs (event recordings), automatically scanning and interpreting paper documentation for a indexing to a digital file cabinet or library. All these problems highlight the need for inferences which take into account the needs of the consumer of the automatically interpreted information.
Lamp design for appearance and performance: As Steve Shafer notes in this volume, most computer vision work is sharply hampered by incomplete methods of predicting visual appearance. I am working on exactly the same problem of how to predict appearance with photometric accuracy, but with a focus on general light source models, lighting inter reflections, and computation. His work is primarily focused in direct reflection. A second aspect of this work is in automating the design of headlamps and taillamps on cars.
Action perception: This is the problem of setting up a camera and a loudspeaker and having the machine narrate, in a natural language, on what is in its view. I believe the synthesis of natural language and computer vision results in deep simplifications which makes such a "story teller" possible. This involves model driven vision where the models are basically action concepts grounded in a given natural language.
|The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.|
Contact Us | Update Instructions