Single-image facial expression recognition using deep 3D re-centralization

Facial expression recognition (FER) aims to encode expression information from faces. Previous studies often hold the assumption that human subjects should properly face the camera. Such a laboratory-controlled condition, however, is too rigid for in-wide applications. To tackle this issue, we propose a single image facial expression recognition method that is robust to face orientation and light conditions. We achieved this by proposing a novel face re-centralization method by reconstructing a 3D face model from a single image. We then propose a novel end-to-end deep neural network that utilizes both re-centralized 3D model and landmarks for FER task. A comprehensive evaluation on three real-world datasets illustrates that the proposed model outperforms the state-of-the-art techniques in both large-scale and small-scale datasets. The superiority of our model on effectiveness and robustness is also demonstrated in both laboratory conditions and wild images.