A Network Model for 3D Object Learning and Dynamic Recognition of Transformed Views

Abstract

This paper proposes a network model that learns multiple views of 3D objects and adaptively recognizes their transformed images. The model achieves view invariance by training bidirectional networks using examples of object views. When novel, transformed images of the objects are then given to the networks, the model performs a dynamic recognition process that iterates between estimating the view using the trained networks and aligning the estimated view with the input view. Computer experiments using gray-level images of 3D objects demonstrate the flexibility of the proposed model.