Real-time Accurate Facial Feature Corner Localization by Rule-Based Selection

We propose a coarse-to-fine approach for localizing eye and lip corners, which is accurate and robust, and can be executed in real-time. Given an image, we first detect the face and the rough initial positions of eyes and lip in the coarsest level. In the middle level, SUSAN corner detection is applied to obtain candidate corners, from which we select a best one as the facial feature corner. Finally, in the finest level, the corner is further refined by normalized cross-correlation (NCC) template matching. We have tested our algorithm on the database containing over 5000 faces with various poses, expressions, occlusions and light conditions. The experimental results show that our method can successfully localize the facial feature corners with more than 20FPS, which is very appealing.