CS547 Human-Computer Interaction Seminar (Seminar on People, Computers, and Design)
Fridays 12:30-1:50 · Gates B01 · Open to the public- 20 years of speakers
- By year
- By speaker
- Videos: iTunesU · YouTube
October 3, 1997
Retrieving images from very large collections using image content as a key is becoming an important problem. The UC Berkeley digital library group has adopted a particular perspective on the problem--namely, that users are primarily interested in scenes containing particular objects or configurations of objects. To make this possible, we believe the key aspects are (a) grouping image data into regions corresponding to objects, or parts of objects (b) recognizing particular configurations of regions as objects based on color, texture, shape or spatial arrangment (c) use learning techniques to assist in the acquisition of common chractersitics of visual categories. We have defined a blob world representation which provides a transition from raw pixel data to a small set of localized coherent regions in color and texture space. Users can construct queries using the blob-world represntation; automatic machine learning techniques can use it to acquire visual categories such as tigers, eagles and airplanes based on a set of examples. We have also demonstrated the use of spatial arrangements of regions (a ``body plan'') to learn and recognize instances of humans and horses from a very large collection of images. |
|