Graphics and Media Lab

Automatic age estimation from face images is an important and actual problem due to its practical significance. Among all existing methods, the best results are achieved by methods based on machine learning algorithms. In such algorithms training set size has a strong influence on error rate. In this paper we propose a way to increase training set size by adding unlabeled pairs of photos of the same person. The only application restriction is that photos within each pair must be taken with insignificant time difference so that the person?s age on both photos is approximately equal. Most face recognition databases meet this condition and enable age estimation rate improvement.