Many CAD-based recognition systems have relied on accurate pose estimation and back-projection in order to verify weak correspondences between simple image and model features. This coupling of recognition and localization requires object models which capture the exact geometry of the object, precluding the recognilion of generic objects in less restricted domains. In this paper, we synthesize a new approach to 3-0 object shape recovery which decouples the processes of recognizing and localizing objects. We first use qualitative shape recovery techniques to recognize objects. Zf and only if detailed shape or pose is required, we then use the recovered qualitative shape to provide strong fitting constraints on physics-based deformable model recovery techniques.