3D Face Reconstruction for Face Recognition and Pose Estimation
Introduction Face pose estimation is an important cue of where the person is directing his or her attention, and thus is useful in human-computer interaction. As classical 2D face recognition methods are sensitive to pose variation, extracting face pose parameters is necessary for further face analysis and recognition.
Research Highlights Different from traditional face reconstruction methods using multiple-view images or learning from a sample database, a simple model-based method using a single near infrared (NIR) frontal image is proposed. With the facial features extracted by Active Shape Model (ASM), I devise a multistep data interpolation algorithm based on Radial Basis Function (RBF) to deform the generic face model. A novel Narrow Band Level Set method without re-initialization is adopted to extract face contour while being fast, which is then used to realistically synthesize face images and as geometric features for face recognition. With the reconstructed 3D face, I present a computation-efficient algorithm via Linear Regression to estimate face pose, which solves the pose parameters between features on the given image with arbitrary pose and those on the synthesized frontal image.
*This project is belong to a larger project: near-infrared face recognition. My work here can facilitate face recognition in three aspects: (1) synthesize virtual face images by the reconstructed 3D model, so as to enlarge training samples and have more samples of a person for recognition; (2) extract face contour by narrow band level set to directly assist recognition; (3) estimate face pose before recognition as a preprocessing step.