Learning View Generalization Functions

TL;DR

This paper introduces a novel framework for 3D object recognition using view-generalization functions, enabling Bayes-optimal recognition directly from limited 2D views without explicit model building.

Contribution

It proposes the view-generalization function framework, unifying existing methods and demonstrating practical learning from few training examples for improved object recognition.

Findings

View-generalization functions can be learned from limited data.

The approach unifies eigenspace and view combination methods.

Experimental results on simulated and real data show effectiveness.

Abstract

Learning object models from views in 3D visual object recognition is usually formulated either as a function approximation problem of a function describing the view-manifold of an object, or as that of learning a class-conditional density. This paper describes an alternative framework for learning in visual object recognition, that of learning the view-generalization function. Using the view-generalization function, an observer can perform Bayes-optimal 3D object recognition given one or more 2D training views directly, without the need for a separate model acquisition step. The paper shows that view generalization functions can be computationally practical by restating two widely-used methods, the eigenspace and linear combination of views approaches, in a view generalization framework. The paper relates the approach to recent methods for object recognition based on non-uniform blurring. The paper presents results both on simulated 3D ``paperclip'' objects and real-world images from the COIL-100 database showing that useful view-generalization functions can be realistically be learned from a comparatively small number of training examples.