Facial Landmark Detection Via Multi-task Feature Selection And Self-adapted Model

Facial landmark detection technology is a hot topic in the field of Computer Vision and Pattern Recognition. It has received more and more attention from public because of its wide application in reality. Facial landmark detection focus on detecting points of face parts we are interested in under natural conditions and changes on faces. Facial landmark detection aims to handle illumination, occlusion, pose and expression change problems, and locate face shape on the face appearance which has both linear and non-linear changes. Nowadays, deep model methods and cascade regression methods are most popular for facial landmark detection. Cascade regression methods are faster and more suitable for daily lives. But there are disadvantages for cascade regression methods. Initialization error is fatal for it, and it's not robust on solving non-linear problems. Thanks to Supervised Descent Method SDM, cascade regression methods can solve these problems to some extent. But there are still initialization problems and over fitting problems, because the feature Supervised Descent Method uses is too simple in some circumstances.So we proposed a Multi-task Feature Selection SDM based on the multi-task learning platform, mixing the popular features state-of-art. By multi-task feature selection we can enhance the robustness of Supervised Descent Method and solve the over fitting problem to make the method more accurate when detecting face landmarks. What's more, we also proposed a Self Adapted Initialization Model for supervised methods. By a fast detect of eyes, the initial face shape is suitable for different faces, and reduce fitting errors in regression as a result. We have carried out experiments against popular baselines on efficient databases. The result shows that our approach achieves state-of-art performance. Besides, we have constructed a synthetical facial landmark detection system for experiments on popular methods and databases at present and the functions are to be extended. The system can also be used to do face detection experiment and face annotation manually, which meets the needs for different facial landmark experiments.