Committee of dataset-biased CNNs for real-world application of facial expression recognition

Abstract

A built-in environment in a dataset plays an important role to decide the performance of a classifier. Up until recently, many facial expression recognition algorithms have competed their performances on a benchmark dataset. However, here one question arises. Does a classifier best in a single benchmark dataset work really better in the real-world environment? To design a classifier working robustly in the real-world, we present an Environment-diversified Network(EdNet). EdNet is a committee of a diverse dataset-biased members which share the feature extraction layers, and 90 of dataset-biased members were trained on 15 blended datasets. Rather than beating the state-of-the-art accuracy on the benchmark dataset, we focused on reducing the accuracy loss of a classifier under unfamiliar environment which the classifier have not been trained on. Finally, we confirmed that EdNet can achieve outperforming cross-dataset generalization by having diversified dataset-biased members.