Siamese PointNet: 3D Head Pose Estimation with Local Feature Descriptor

Head pose estimation is an important part of the field face analysis technology. It can be applied to driver attention monitoring, passenger effective information screening, etc. However, illumination changes and partial occlusion interfere with task, due non-stationary characteristic head change process, normal regression networks are unable achieve very accurate results on large-scale synthetic training data. To address above problems, a Siamese network based 3D point clouds was proposed, which adopts share weight similar samples constrain process pose’s angles; meanwhile, local feature descriptor introduced describe geometric features objects. In order verify performance our method, we conducted experiments two public datasets: Biwi Kinect Pose dataset Pandora. The show that compared latest methods, standard deviation reduced by 0.4, mean error 0.1; also maintained good real-time performance.