Study On Cross-dimension Facial Landmark Localization

With the development of computer vision technology in recent years,there is an increasing applications of facial landmark localization,e.g.face unlock,face effect in video,and 3D face construction.However,there are still some issues for landmark localization caused by scenarios variation.2D and 3DA-2D(2D Projections of 3D Annotations)landmarks are two different kinds of facial landmarks for 2D images.2D landmarks are used to describe face contour while3DA-2D landmarks are used to describe 3D face shape.Obviously,the two landmarks have different application scenarios.In this paper,we propose a fast cross-dimension conversion method between 2D landmarks and 3DA-2D landmarks based on cascade regression,which provide a solution for landmarks correlation among various scenarios.In consideration of effects of face scale for face alignment,we propose an adaptive face scaling method which avoid image normalization.Next,given undesirable performance of traditional cascade regression methods under massive training data or noisy data,we present MBCR method,i.e.Mini-Batch Cascade Regression,to overcome these issues.Compared with traditional cascade regression methods,MBCR realizes an online learning and model fine-tune with massive training data,and it also has a better robustness.Since the initial face shape has a considerable influence on 2D landmarks localization,we propose a robust initial face shape generation method.The method can work with different face detection frames and guarantee a constant initial shape for different faces.Then,to find out the correlation of facial landmarks among various scenarios,we study the fast conversion between 2D and 3DA-2D landmarks,we called it Cross-Dimension Annotations Conversion.In this paper,both of the 2D landmarks localization and CDAC are realized by MBCR.Finally,we conduct experiments to explore the performance of MBCR.The result shows that MBCR outperforms other state-of-the-art methods on accuracy in 2D and 3DA-2D landmarks localization.In addition,for CDAC between 2D and 3DA-2D landmarks,we obtain a 110 fps conversion with 3.4GHz CPU.