Abstract

We describe our approach towards building an efficient predictive model todetect emotions for a group of people in an image. We have proposed thattraining a Convolutional Neural Network (CNN) model on the emotion heatmapsextracted from the image, outperforms a CNN model trained entirely on the rawimages. The comparison of the models have been done on a recently publisheddataset of Emotion Recognition in the Wild (EmotiW) challenge, 2017. Theproposed method achieved validation accuracy of 55.23% which is 2.44% above thebaseline accuracy, provided by the EmotiW organizers.

 

Quick Read (beta)