Automatic emotion classification vs. human perception: Comparing machine performance to the human benchmark

Emotion classification is performed by humans in all kinds of situations. However, perception tests, in the literature, show that humans do not perform perfectly even when classifying prototypically performed expressions. The fact that humans commit errors in this task, demonstrates that higher performance accuracies do not directly imply more realistic comprehension of the emotions´ nature. Therefore, in this study we do not aim for perfect recognition performances, but rather analyze the influence of the derived features to the overall classification results and compare results to human perception tests. For this purpose, our experimental results, achieved with multi-classifier multi-class support vector machines, combining eight separate feature sets, are based on standard datasets. The results are compared, using confusion matrices, with the human perception capabilities, yielding similar accuracies.