Learning Visual Ideals

We address the problem of describing, recognizing, and learning generic, free-form objects in real-world scenes. We introduce a hybrid appearance-based approach, IDEAL, where objects are encoded as a loose collections of parts and the relations between them. The key features of this new approach are the structural part decomposition combining multi-scale wavelet segmentation and hierarchical blobs, and learning to recognize generic object categories, exhibiting large intra-class variability , from real examples with automatic model acquisition. Then, my dear friend, must not the law-giver also know how to embody in the sounds and syllables that name which is tted by nature for each object? Must he not make and give all his names with his eye xed upon the absolute or ideal name, if he is to be an authoritative giver of names? And if diierent lawgivers do not embody it in the same syllables, we must not forget this ideal name on that account; for diierent smiths do not embody the form in the same iron, though making the same instrument for the same purpose, but so long as they reproduce the same ideal, though it be in diierent iron, still the instrument is as it should be, whether it be made here or in foreign lands, is it not? { Socrates in Plato's Cratylus x389e As Socrates explained to Hermogenes, we carry a conceptual ideal of many objects in our thoughts. Indeed the concept of ideals are embedded in our very languages, witness the almost universal usage of deenite and indeenite articles. When we say \a chair", we envision in a sense the ideal form of a chair. One does not recall an actual chair, but instead a functional description of what a chair does combined with a relational-structural description of the parts which make up a chair. In our system, IDEALL2,3], we attempt to learn from example objects the structural parts and the relations between them which when recognized in a previously unseen object, enable the system to probabilistically classify it amongst its known classes. 1 3-D Object Recognition Recognizing three-dimensional objects under diierent viewing and lighting conditions is a traditional problem in computer vision. The diiculty of the problem