Deep Modeling of Human Age Guesses for Apparent Age Estimation

In this paper we propose a unique deep learning formulation of the apparent age estimation problem, using the APPA-Real dataset. APPA-Real is a dataset containing 7, 591 face images, where each image is labeled by a set of approximately 38 guesses of the facial age. All guesses are collected from human labelers. In our approach, we first generate per-image label distributions from the human guesses, and then learn label distributions with convolutional neural networks and the KL-divergence loss function. We provide comparisons to models trained with other objective functions. We achieve state-of-the-art results for apparent age estimation on the APPA-Real dataset with a mean absolute error of 3.688, outperforming other methods using the same dataset.