No abstract

Abstract

In this paper we study some problems important for large-scale human age estimation. First, we study age estimation performance under variations across race and gender. Through a large number of age estimation experiments, significant differences are observed for age estimation between "no crossing" and "crossing." Our study discovers that crossing race and gender can result in significant error increases for age estimation. This finding provides a guide for age estimation in practice, especially for cross-database experiments. Second, we propose a complete framework for crossing race and gender age estimation, based on our findings. Third, age estimation is performed on the large database of MORPH-II with more than 55,000 images. A small MAE of 4.45 years is obtained based on our proposed framework, which is much smaller than a recently reported MAE of 8.60 years on MORPH-II.

No abstract

No abstract