We present a solution for automatic extraction of contour models out of polygonal CAD models. Such contour models can be used for markerless initialization and tracking purposes. To create a contour model we synthesize different views of the object using its CAD model. During the view generation we alternate the camera pose as well as light conditions in order to extract the most stable contours in terms of illumination invariance and view independence. We project the extracted 2D edges back into the 3D space and accumulate for every 3D point statistics over different views describing its visibility and stability under different illumination conditions. After filtering out all 3D points with probability below a certain threshold value we use the Euclidean Minimum Spanning Tree algorithm to get the connected contours out of the 3D point cloud. The result is a B-Spline or VRML representation of the most stable contours.

Abstract Not everything is perceived as it is provided by the environment. Depending on focus and attention perception can vary and therefore also the knowledge about the world. Virtual humans are sensing the virtual world, storing knowledge and using it to perform tasks. This paper describes our approach to model perceiving, storing and forgetting knowledge as the main regulation of tasks. We use different forms and levels of knowledge which can be independently adapted to different personalities and situations by combining computer graphics methods with psychological models. Copyright © 2005 John Wiley & Sons, Ltd.