Detection And Labeling Of 3D Face Feature Points

The automatic detection of 3D facial feature points has a very important role in face recognition,tracking, modeling, and expression analysis. Existing studies generally assumed that characteristic points is the extreme of a function on three-dimensional surface,but this assumption is only suitable for the nose and outstanding points, such as the inner eyes, not for under prominent feature points. Therefore, on the basis of analyzing and summarizing previous previous work, we propose a new method which can locate the weak local shape of 3D facial feature points. Details are as follows:(1)We proposed a feature detection method based on multiple description matching distribution and combination. We use nine descriptors, like the maximum principal curvature, minimum principal curvature, Gaussian curvature, mean curvature, shape index and so on. We combinate these descriptors by using combination rules learned in the process of training. Experiments shows that single local descriptor reponses of some feature points is weak, but after combinating these descriptors by using this method,information of all feature points has become clearer, this makes detection and location easier.(2) Two methods are learned to combinate descriptors in the process of training.Firstly, call 14 hand-labeled points as reference points,and calculate nine descriptors on all mesh points; secondly, observe and estimate the probability distribution parameters of the, then all points on the mesh are calculated using the same parameters and every points are calculated likelihoood with the responding reference ponit, in other words,every point are calculated the ratio of the maximum probability distribution density of the corresponding reference point estimation to normalized, then obtaining a matching score distribution; one more, using the linear discriminant analysis and AdaBoost learning respectively merge rules to generate a local shape score, Makethe reference point in the neighborhood of local shape high score and the the neighborhood within the local shape low score. Combination rules are probability distribution parameters and weights for combinating. Experiments shows that results produced by these two methods are similar.(3) We proposed a marking method based on RANSAC registration and projection. The input are extracted directly from the local shape scores of local minima as candidate markers. Then use a rigid model based on adaptive scale random sampling algorithm consistency registration candidate marker point set and the reference model, selected to match the reference point is the most points. Finally, calculate the projection distance between the point on the calculation and the selected point on the model, selecting the nearest point as a marker.