This project investigates vision features and their intelligent application on known CAD models for pose estimation. The advantage of combining CAD models and vision features lies in the known 3D geometry of the CAD model. Exemplary vision features are e.g. SIFT features, Randomized Trees, SURF features and other well known methods. The main goal is to use such an extended CAD model for pose estimation and for tracking initialization.
Natural 3D Markers (N3M)
A first step towards vision features incorporating 3D-CAD models was made by creating Natural 3D Markers (N3Ms). N3Ms are minimal subsets of vision features which are learned during a training step from a 3D-CAD model and which are carefully designed to guarantee fast detection and quasi optimal pose estimation. N3Ms lead to high invariance to partial occlusion, high invariance to cluttered backgrounds, fast outlier elimination and the absence of ill-conditioned point configurations (e.g. collinear points) in the detection and pose estimation process. The results were published at ICCV 2007 in Rio de Janeiro.
The detection and pose estimations runs with about 15fps.