Recognition by combinations of model views: Alignment and invariance

Abstract

A scheme for recognition of 3D objects from single 2D images is introduced. An object is modeled in this scheme by a small set of its views with the correspondence between the views. Novel views of the object are obtained by linearly combining the model views. The scheme accurately handles rigid objects under weak-perspective projection, and it is extended to handle rigid objects with smooth bounding surfaces and articulated objects. Unlike in other schemes, explicit 3D representations of the objects are not used. The presented scheme can be used both under an alignment framework and as a means for deriving object-specific invariant functions for indexing. Under an alignment framework, given a model and an image, the coefficients of the linear combination that aligns the model with the image need to be recovered. A small number of points in the image and their corresponding points in the model can be used for this purpose, or a search can be conducted in the space of possible coefficients. Alternatively, the scheme can be used to derive functions that are invariant to viewpoint changes of a specific object. A number of such functions are derived in this paper.