Computer Vision Research Group

Object Detection and Categorization in Natural Scenes

While imitating human performance, cognitive technical systems should have an automatic visual component. In order to be useful such a component should be able to recognize many different object classes, as usually many objects are visible in an observed scene, and as goals of the system may vary strongly. This is only possible with generic object detection methods which can be trained to detect specific objects in a short time period. Presently, the most promising approach to achieve this goal is the use of unsupervised object categorization. The procedure is started with the extraction of shift-, rotation- and scale-invariant features at the location of prominent points in the image. The extracted parameters include a description of the location relative to other features.

Such graph-based information is utilized to identify previously unknown object categories. Research efforts are directed to enable systems to recognize new object classes with no or only few training samples acknowledging the fact that humans and particularly children have the ability to recognize new object classes and learn them with only one or two examples. This goal could possibly be reached by sharing the same feature types among several object classes. Recent research results have shown that then the numbers of features per object class necessary for identification increase, but that far less feature types are necessary to run the complete system. This system behavior is based on an implicit use of prior experiences and could probably be extended to incremental learning abilities.

To identify objects robustly also context information has to be used, i.e. the complete visible scene has to be considered to belong to the object. Examples from human daily life proving this requirement are manifold. To make the approach practically operational it should be included in an interactive system. Human operators could show training samples; the system would automatically improve its recognition performance by training itself, e.g. using web-based data mining procedures. The research project proposed would be followed in co-operation with participating machine learning research groups.(see related publications)

(Joint work with Prof Hellwich group, Computer Vision & Remote Sensing, TU Berlin, Germany, www.cv.tu-berlin.de )



Recent News
© 2010 Department of Electrical Engineering, CIIT Lahore  |  Terms and Conditions