Cross-database evaluation for facial expression recognition

We present a system for facial expression recognition that is evaluated on multiple databases. Automated facial expression recognition systems face a number of characteristic challenges. Firstly, obtaining natural training data is difficult, especially for facial configurations expressing emotions like sadness or fear. Therefore, publicly available databases consist of acted facial expressions and are biased by the authors' design decisions. Secondly, evaluating trained algorithms towards real-world behavior is challenging, again due to the artificial conditions in available image data. To tackle these challenges and since our goal is to train classifiers for an online system, we use several databases in our evaluation. Comparing classifiers across data-bases determines the classifiers capability to generalize more reliable than traditional self-classification.