Real-time Convolutional Neural Networks for Emotion and Gender\n Classification

Abstract

In this paper we propose an implement a general convolutional neural network\n(CNN) building framework for designing real-time CNNs. We validate our models\nby creating a real-time vision system which accomplishes the tasks of face\ndetection, gender classification and emotion classification simultaneously in\none blended step using our proposed CNN architecture. After presenting the\ndetails of the training procedure setup we proceed to evaluate on standard\nbenchmark sets. We report accuracies of 96% in the IMDB gender dataset and 66%\nin the FER-2013 emotion dataset. Along with this we also introduced the very\nrecent real-time enabled guided back-propagation visualization technique.\nGuided back-propagation uncovers the dynamics of the weight changes and\nevaluates the learned features. We argue that the careful implementation of\nmodern CNN architectures, the use of the current regularization methods and the\nvisualization of previously hidden features are necessary in order to reduce\nthe gap between slow performances and real-time architectures. Our system has\nbeen validated by its deployment on a Care-O-bot 3 robot used during\nRoboCup@Home competitions. All our code, demos and pre-trained architectures\nhave been released under an open-source license in our public repository.\n