Towards East Asian Facial Expression Recognition in the Real World: A New Database and Deep Recognition Baseline.

In recent years, the focus of facial expression recognition (FER) has gradually shifted from laboratory settings to challenging natural scenes. This requires a great deal of real-world facial expression data. However, most existing real-world databases are based on European-American cultures, and only one is for Asian cultures. This is mainly because the data on European-American expressions are more readily accessed and publicly available online. Owing to the diversity of huge data, FER in European-American cultures has recently developed rapidly. In contrast, the development of FER in Asian cultures is limited by the data. To narrow this gap, we construct a challenging real-world East Asian facial expression (EAFE) database, which contains 10,000 images collected from 113 Chinese, Japanese, and Korean movies and five search engines. We apply three neural network baselines including VGG-16, ResNet-50, and Inception-V3 to classify the images in EAFE. Then, we conduct two sets of experiments to find the optimal learning rate schedule and loss function. Finally, by training with the cosine learning rate schedule and island loss, ResNet-50 can achieve the best accuracy of 80.53% on the testing set, proving that the database is challenging. In addition, we used the Microsoft Cognitive Face API to extract facial attributes in EAFE, so that the database can also be used for facial recognition and attribute analysis. The release of the EAFE can encourage more research on Asian FER in natural scenes and can also promote the development of FER in cross-cultural domains.