Integrating Qualitative and Quantitative Object Representations in the Recovery and Tracking of 3-D Shape

Data-driven models such as active contours in 2-D and deformable surfaces in 3-D have become prevalent in the computer vision community, particularly in the areas of shape tracking and shape recovery. They provide an important alternative to typical model-based recognition and tracking approaches that assume knowledge of the exact geometry of the object. Despite the power of these approaches, data-driven models often encode too little model information. As a consequence, active 3-D model recovery schemes often require manual segmentation or good model initialization, and active contour trackers have been able to track only an object's translation in the image. To overcome these problems requires bridging the representational gap between overconstrained geometric models and underconstrained active models. In previous work, we introduced a qualitative object representation integrating object-centered and viewer-centered models. In this paper, we rst show how this representation provides the missing constraints on the recovery of quantitative 3-D deformable models from 2-D images. We then show how this same representation provides the missing constraints needed to qualitatively track an object's rotation in depth or to quantitatively track an object's pose.