VIVARIUM Seminar - Finding Objects with Dynamically Deformable Models
January 24, 2008
Location: TASC I - 9204
Presenter: Klaus D. Toennies
Dynamically deformable models represent expected object shape or expected object shape attributes. A representation of expected shape may be used for finding object instances provided that they sufficiently discriminate within- and inter-class object variation. No feature extraction step is needed since model adaptation is carried out directly on the data. We present methods and results from recent experiments on applying different model representations for finding objects in data. The first case, the detection of lymph nodes in CT, is relatively simple because of a restricted variation of object shape. In this case, model deformation can be used to indicate detection of an object. The strategy is not successful, however, if object variation is too large or if objects may be distorted or partially hidden. Hence, a prototypical FEM was extended to represent object-specifying and class-discriminating features on different levels. Different to an earlier approach, within- and between-level interaction was defined in a uniform fashion. Structural adaptation may influence morphological adaptation and vice versa.
Object search simultaneously adapted model instances to object- and class-specific features. The strategy will be demonstrated for detecting Heschl’s gyrus in Flatmaps from segmented MRI and for classifying ant species from pictures of ants in a data base. Results show adequateness of the components of the representation and a very robust behaviour in the presence of large deformations or hidden object parts.
Klaus D. Toennies received his MSc and PhD in Computer Science in 1983 and 1987 from the Technical University Berlin. From 1987 to 1989, he was Research Assistant Professor at the Medical Image Processing Group of the University of Pennsylvania. From 1989 to 1995, he was Assistant Professor at the Computer Graphics Group at the Technical University Berlin. From 1996 to 1998, he was leading the Image Processing Group in the Department of Radiology at the University Berne, Switzerland. Since 1998, he is Professor (Image Processing and Image Understanding) in the Department of Computer Science of the University Magdeburg. His current research interests are on the use of shape models for object recognition and object analysis. Applications are mainly related to medical imaging.