부산대학교 도서관

This paper demonstrates a system for distinguishing between real and posed smiles from 25 observers’ accelerometer data. The data are gathered using Empatica E4 wristband when observers were watching four kinds of smiles, namely single image (SI), single video (SV), paired images (PI) and paired videos (PV). A two sample K-S test shows that the accelerometer data, which are recorded when observers are watching a single video (SV) of displayers’ smiles, is more reliable compared to other conditions. We computed performance of the accelerometer-based smile detection process using decision tree, logistic regression, K-nearest neighbor, stochastic gradient descent, naïve bayes, and random forest classifiers. Based on the selected sub-datasets, it is found that the decision tree outperforms all the other classifiers, with 80% classification accuracy on our two classes of smiles (real vs. posed). This high accuracy indicates the possibility that the accelerometer data can be used for distinguishing between human real and posed smiles.