Ellipsoidal Bias in Learning Appearance-Based Recognition Functions

Abstract

We present an approach for learning appearance-based recognition functions, whose novelty is the sparseness of necessary training views, the exploitation of constraints between the views, and a special treatment of discriminative views. These characteristics reflect the trade-off between efficiency, invariance, and discriminability of recognition functions. The technological foundation for making adequate compromises is a combined use of principal component analysis (PCA) and Gaussian basis function networks (GBFN). In contrast to usual applications we utilize PCA for an ellipsoidal interpolation (instead of approximation) of a small set of seed views. The ellipsoid enforces several biases which are useful for regularizing the process of learning. In order to control the discriminability between target and counter objects the coarse manifold must be fine-tuned locally. This is obtained by dynamically installing weighted Gaussian basis functions for discriminative views. Using this approach, recognition functions can be learned for objects under varying viewing angle and/or distance. Experiments in numerous real-world applications showed impressive recognition rates.