| Facial Landmark Detection is a technology that machine automatically detects points with specific meanings on human face according to face information.It has been widely applied in many aspects,such as Face Recognition,Face Segmentation,3D Face Reconstruction,Posture Estimation and etc.This technology can significantly improve the efficiency and accuracy of these tasks.After years of development,the tasks related to human faces have become more and more complex.The focus of Facial Landmark Detection has gradually shifted from improving the speed and accuracy of detection to improving the difficulty of detection,which is called Dense Facial Landmark Detection.At the same time,researchers' demand for data sets is becoming more and more urgent,because databases are always of great significance to researchers to achieve a satisfactory model.Annotating dense landmarks is a very tedious work due to two challenges.(a)Not every facial point has clear definition.Some of them distribute uniformly along the contour.Their labeled positions are determined by subjective judgment of the annotators,so that the quality of the annotation is normally poor.(b)Adjusting facial points one by one is time-consuming.The workload will increase dramatically when there are more points.During this resarch,the key points of face are studied and analyzed in depth,and solutions to the above problems are put forward.The main contributions of this paper are as follows:(1)A new definition of dense key points is given.A total of 84 points are defined on the face.At present,the widely used definition of key points of dense face contains 68 points,but it lacks the definition of the two sides of nose and the lower half of eyebrow,and the key points of eyes are sparse.This paper adds definitions of key points for eyebrows,noses and eyes,so that they can accomplish more complex tasks.(2)The AFW,HELEN,IBUG,LFPW and some Multi-PIE images were re-labeled,and some images searched from the website were added.The Dense Landmark Localization Database(DLL)which contains 39198 pictures was constructed and each face image was labeled with 84 facial key points.These face pictures included posture,expression,illumination and occlusion changes,and pictures in laboratory environments and uncontrolled scenes were both considered.It is worth mentioning that the data set will be published and provided to researchers for free use.(3)The Supervised Descent Method and Convolution Pose Machine algorithm are evaluated on the data set.The experimental results are summarized and analyzed.A face key point detection method based on datum loss function is proposed,which achieves higher detection accuracy on difficult samples.(4)A semi-automatic facial key point annotation tool is designed,which combines registration algorithm and three-dimensional face reconstruction.It makes full use of the structural information of the face itself to reduce the dependence on the subjective judgment of the annotator.At the same time,the tool can semi-automatically move the key points to the potential target location,so that the annotator can annotate dense facial key points with fewer clicks.It effectively improves the efficiency and accuracy of tagging.Overall,this paper provides a data set labeled with dense face key points,which covers a wide range of pictures and has high labeling accuracy.To a certain extent,it alleviates the demand of deep learning for data set.On the other hand,through the study of key point location and face structure,a semi-automatic tagging tool is proposed,which improves the efficiency and accuracy of annotation work.The DLL data set constructed in this paper and the semi-automatic facial key point annotation tool will be open source for researchers to use for free. |
PDF Full Text Request