An Academic Emotion Database and the Baseline Evaluation

Academic emotion is one of the important factors affecting learning effect, which exists on various learning situations. Realizing its automatical inference has a significant meaning for educational field. Facial expression recognition (FER) is always selected as its basis. Nevertheless, the preconditions for FER include a high quality database with sufficient training samples, which is not available currently. In this paper, the previously developed facial expression databases are reviewed and a novel database is established. It includes spontaneous facial expressions of five common academic emotions. 30,184 images from 82 students are included, which are collected in a real online learning environment. Both participants themselves and trained coders are invited to label the samples. The reliability of the encoding is validated by Fleiss's kappa 0.867 and Cohen's kappa 0.943, which is good. For baseline evaluation, two kinds of recognition algorithms are employed with 10-fold cross-validation over participants. One is the combination of different feature extraction technologies and classifiers and the other is Convolution Neural Network (CNN). The best result is obtained by CNN with 0.877 accuracy. Such a database will certainly accelerate the application of affective computing in educational field.