With a shift in interest toward dynamic expressions, numerous corpora of dynamic facial stimuli have been developed over the past two decades. The present research aimed to test existing sets of dynamic facial expressions (published between 2000 and 2015) in a cross-corpus validation effort. For this, 14 dynamic databases were selected that featured facial expressions of the basic six emotions (anger, disgust, fear, happiness, sadness, surprise) in posed or spontaneous form. In Study 1, a subset of stimuli from each database (N = 162) were presented to human observers and machine analysis, yielding considerable variance in emotion recognition performance across the databases. Classification accuracy further varied with perceived intensity and naturalness of the displays, with posed expressions being judged more accurately and as intense, but less natural compared to spontaneous ones. Study 2 aimed for a full validation of the 14 databases by subjecting the entire stimulus set (N = 3812) to machine analysis. A FACS-based Action Unit (AU) analysis revealed that facial AU configurations were more prototypical in posed than spontaneous expressions. The prototypicality of an expression in turn predicted emotion classification accuracy, with higher performance observed for more prototypical facial behavior. Furthermore, technical features of each database (i.e., duration, face box size, head rotation, and motion) had a significant impact on recognition accuracy. Together, the findings suggest that existing databases vary in their ability to signal specific emotions, thereby facing a trade-off between realism and ecological validity on the one end, and expression uniformity and comparability on the other.
In the present paper, we focus on the empirical evaluation by measuring and comparing emotion recognition indices across individual databases. For this purpose, we collected data from human observers and conducted automated facial expression analysis with a software tool called FACET (iMotions). FACET has been used widely, thereby demonstrating superior levels of emotion classification in recent cross-classifier comparisons (Stöckli et al., 2018; Dupré et al., 2020).
