Eye Center Guided Constrained Local Model for Landmark Localization in Facial Image

Landmark localization is a very important step for many face-related computer vision applications. Compare to the holistic approaches (e.g. AAMs), constrained local models (CLMs) shows good performance for landmark localization in non-rigid facial images. But these methods are always limited by the initialization. This paper proposed an eye center guided constrained local model where the initialization is performed by mean face shape taking eyes as references. First, we have adopted a hybrid eye detector method to find both the eye centers and then mean face shape is placed on the basis of orientation and distance of two eye centers. Moreover, we have analyzed our models with some descriptors to find the best descriptor to represent our model. The proposed CLM approach has been tested on AR and Multi-PIE databases with 130 and 68 landmarks respectively. The experimental results suggest that our proposed method has achieved improved performance as compared to existing methods.