Apparent Age Estimation with Relational Networks

Apparent age estimation is a newly proposed and under-studied problem of predicting the age that someone "looks" rather than their actual age. It has applications in many areas within the beauty industry[3]. Methods based on convolutional neural networks (CNNs) have proved to be state-of-the-art on the few datasets used to benchmark this task. However, such CNNs typically collapse spatial information via a Global Average Pooling operation. They do not perform any explicit treatment of spatial relationships of the higher-level features which emerge in the later stages of the network and which may correspond to facial parts or blemishes that are characteristic of age. In this paper, we consider a newly proposed CNN module called relational networks that explicitly capture spatial relationships. We hypothesize that we can estimate age better by learning such relationships in the final set of CNN feature maps where spatial information is still retained. Experiments were conducted on both ChaLearn LAP 2015 and 2016 datasets [6], [7] showing that on average, there is a 3.53% improvement on Mean Absolute Error and 3.31% improvement on ε-error when compared to the baseline. A test was also calculated to show that the improvement is statistically significant.