Robust pose estimation for arbitrary objects in complex scenes

Viewer-centered estimation of the pose of a three dimensional object has two main advantages: No explicit models are needed and error-prone corner detection is not necessary. Eigenspace methods have been successful in pose estimation especially for faces. However, most eigenspace-based algorithms fail if the images are corrupted, e. g. if the object is occluded, the background differs from the training images or the image is geometrically transformed. EigenTracking by Black and Jepson uses robust estimation to find the correct pose. We show that performance degrades for objects whose silhouette changes greatly with 3D rotation. To solve this problem we introduce masks that adapt to the estimated object pose. To this end we used hierarchical eigenspaces containing both the appearance and mask descriptions. We illustrate the improvement in pose estimation precision for some typical objects.