Beyond Wisdom of Crowds: Deep Uncertainty Coding for Apparent Age Estimation

Conventional algorithms for apparent age estimation make a strong assumption that the wisdom of crowds (i.e. the mean value of inconsistently annotated age labels) is superior to any individual. In this paper, we go beyond such an assumption and instead of averaging to remove label uncertainty we cope with label uncertainty in training by introducing a new concept of uncertainty coding. Uncertainty coding encodes multiple unary apparent age labels into a binary vector indicating age range (uncertainty) in the target space. We traverse through convolutional features and architecture to learn deep uncertainty mapping, whose predicted codes as mid-level features are fed into the second layer regressor that outputs a target prediction. We conduct experiments on the ChaLearn and FG-NET benchmarks to demonstrate superior performance of the proposed deep uncertainty coding to the state-of-the-arts.