Simultaneous Recognition and Homography Extraction of Local Patches with a Simple Linear Classifier

CAMP | Computer Aided Medical Procedures | http://campar.cs.tum.edu | Technische Universitat Munchen | TUM References [1] Hinterstoisser, S. et al.: Online Learning of Patch Perspective Rectification for Efficient Object Detection, CVPR 2008, Anchorage, Alaska, USA. [2] Ozuysal, M. et al.: Fast Keypoint Recognition in Ten Lines of Code, CVPR 2007, Minneapolis, Minnesota, USA. [3] Jurie, F. et al.: Hyperplane approximation for template matching, PAMI, 2002 Problem •Previous approaches (Leopar1 [Same authors,CVPR'08]) showed that we can efficiently estimate the 3D pose of a poorly textured object by learning the patch appearance. Leopar1 is performed in 4 steps: (a) Pre-classify feature points with e.g. Ferns2 (b) Rough orientation estimation of the patch with respect to the feature point identity (c) Rectification refinement by applying a Template Matching algorithm3 (d) Outlier removal by simple correlation measure Leopar1 gives much better results concerning the accuracy and repeatability of the pose than affine region detectors but still suffers from a decreasing robustness towards large viewpoint changes. Reasons for that: • Error-prone pre-classification of initial feature points in (a) (by Ferns2) without taking into account the pose of the patch • Rough estimation of the pose of the patch in (b) is limited to orientations