Sequence-based bias analysis of spontaneous facial expression databases

In this paper, cross-corpora evaluations are used to analyze the bias of spontaneous facial expression databases. Local binary pattern and Gabor feature are extracted from difference-image sequences. A Hidden Markov Model is used as the classifier to discriminate arousal (i.e. high versus low) and valence (i.e. positive versus negative) respectively. Four datasets are adopted, including: UT-Dallas, USTC-NVIE, DEAP and MAHNOB. Experimental results indicate that there exists bias among different spontaneous facial expression databases. The bias may reduce the generalization performance of algorithm trained on these databases. The emotion induction stimulus, the variety of subjects, and the segmentation may have caused such a bias.