Face Alignment And Face Pose Estimation Based On Boundary Map And Classification

Face alignment and face pose estimation are two important face tasks,face alignment is to describe the shape of the face,and the face pose estimation is to describe the position of the face relative to the camera,both of which are important features for describing faces.The accurate face alignment algorithm can help estimating the subsequent facial pose.For existing face alignment algorithms,different facial landmarks have different convergence speeds and prediction difficulties.Since all the landmarks are trained equally,the accuracy is limited;for the loss function,the relative distance and direction between the landmarks have always been important factors that are ignored.For existing face pose estimation algorithms,facial shape plays a vital role in pose estimation prediction.Due to the complex face environment,it is difficult to input an ordinary face image to accurately feed the face shape to the network;Face pose estimation is a regression problem.Due to the complexity of the environment and the difficulty of optimization,the accuracy of face pose estimation is low in this situation.In response to the above problems,we propose a two-stage face alignment algorithm,a face pose estimation algorithm based on boundary maps and classification.The main contribution is:1.Face alignment algorithm with landmark clustering and contraint loss function.The main contribution is:(1)for the problems of different convergence speeds and large differences in positioning difficulty in face alignment training.We use unsupervised clustering K-Means to divide the face landmarks based on convergence speed and positioning difficulty,and different clusters of landmarks are processed by different networks.(2)Add a factor that constrains the shape of the face to the loss function to help the network learn the relative information between landmarks and landmarks.We evaluate the proposed algorithm on the public data sets 300W and WFLW,and achieved inter-ocular error of 3.26%and 4.32%respectively.2.The proposed multi-class pose regression algorithm with the face boundary maps and classification.The main contribution is:(1)Add face boundary maps to the input of the regressor to assist feature extraction.(2)Convert the regression problem into a mixed classification and regression problem for constrained optimization.We evaluate algorithm proposed in this article on the public data sets AFLW2000-3D and BIWI,and achieve the average angle errors of 3.89 and 4.54 respectively.