A Probabilistic Approach to Fast and Robust Template Matching and its Application to Object Categorization

This paper presents a new statistic, called probabilistic increment sign correlation (probabilistic ISC), for evaluating similarity between images of objects which have intra-class variation such as individual differences of human faces. The new statistic evaluates similarity between an input image and object classes, whereas most conventional methods, such as normalized cross-correlation, calculate correlation between an input image and a template. The new statistic is defined as a log-likelihood based on probabilities of observing the increment signs. Probabilistic ISC provides two advantages over conventional correlation-based methods: 1) robustness against the intra-class variation because it gives larger weights to stable features which are commonly observed in reference images and 2) robustness against noise and change in illumination. It yields higher performance even if a small number of reference images are given, whereas other methods such as the subspace method and AdaBoost cannot maintain their accuracy. We show these advantages through several experiments of face detection and face orientation estimation