Strong-from-weak model sensor estimation using Voronoi diagrams

Rendering 3D models with photorealistic texture is an emerging problem when dealing with large 3D model databases. Despite the fact that techniques for manual extraction of image patches are well known, the following method aims at extracting patches in an automatic manner, given a CAD model of a scene to be textured. We describe the recognition phase, and suggest a solution employing a strong-from-weak alignments of corner hypothesis in the image space. A 2D Voronoi diagram permits us to extend the number of observations so as to infer a more complex sensor model. Due to multiple detection of images corners, the mapping between model scene and image features may be one-to-many. A solution which finds the best minimal matching error based on an interpretation tree is suggested. The matching error is a global measure that takes into account the sensor parameters and the CAD model uncertainty. The following algorithm was successfully tested on the Radius image and scene model database.