Emotions, and consequently facial expressions, play an essential role in communication-and thus in everyday life. With the increase of human-machine interactions, and more especially of multimedia applications, automatic recognition of facial expressions has emerged as a challenging task, particularly under naturalistic conditions. In the present work, a benchmark is firstly conducted using four open source deep learning solutions on four labeled image datasets. Thanks to an exhaustive analysis based on two distinct, yet complementary approaches, we show how the four models performed depending on the studied emotions. Furthermore, we present a novel metric based on the Euclidean distance between two given emotions (i.e., ground truth and predicted) to better measure the performance of said models in the context of interactive media, where human sensibility needs to be taken into consideration.