Assessing facial age similarity: a framework for evaluating the robustness of different feature sets

Abstract

A framework that can be used for assessing the suitability of different feature vectors in the task of determining the age similarity between a pair of faces is introduced. This framework involves the use of a dataset containing images displaying compounded types of variation along with the use of an ideal dataset, containing pairs of age-separated face images captured under identical imaging conditions. The use of the ideal dataset in conjunction with deliberate introduction of controlled noise, allows the extraction of conclusions related to the robustness of different feature vectors to different types of noise effects. The ultimate aim of this work is the derivation of comprehensive and accurate set of metrics for evaluating the performance of age progression algorithms in order to support comparative age progression evaluations.