A new descriptor of gradients Self-Similarity for smile detection in unconstrained scenarios

Smile detection is a sub-problem of facial expression recognition field, which has attracted more and more interests from researchers because of its wide application market. As for smile detection problem itself, the ‘wild’ unconstrained scenario is more challenging than the laboratory constrained scenario. Therefore, in this paper, we mainly focus on solving smile detection problem in unconstrained scenarios. To this end, a new descriptor, Self-Similarity of Gradients (GSS), is proposed. Inspired by Self-Similarity on Color channels (CSS) feature in pedestrian detection area, GSS can effectively describe the similarities in a HOG feature map, while these similarities are useful and helpful for constructing a high-performance practical smile detector. Moreover, since a smile detector using multiple features and multiple classifiers simultaneously shows superior performance, they are also adopted by us. Finally, experimental results indicate that the combined features (HOG31+GSS+Raw pixel) using AdaBoost with linear Extreme Learning Machines (ELM) achieve improved performance over the state-of-the-arts on the real-world smile dataset (GENKI-4K).