Facial Age Estimation by Total Ordering Preserving Projection

Facial age estimation is one of the unsolved challenging issues in automatic face perception. Previous studies usually formulated it as a classification problem, where each age is regarded as a class, or a regression problem where the age is regarded as a variable spanning in a real-valued interval. In this paper, we propose to formulate this task as an ordinal regression problem. On one hand, the new formulation emphasizes the fact that the age estimation problem is inherently a classification problem (ordinal regression is a special kind of classification task); on the other hand, the new formulation allows to take into account the order information between different ages, which has been ignored by previous classification formulation. We develop the TOPP (Total Ordering Preserving Projection) approach, by identifying the low-dimensional subspace which preserves the ordinal relations to the best, and experiments show that TOPP significantly outperforms state-of-the-art age estimation methods.