PublicaHier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.
Shape optimization of a deformable 3D face model with respect to multiple 2D views
In a face or facial expression recognition system, one of the key requirements is the proper generation of a face model. An approach for reconstructing the 3D face model from a video sequence is presented within this thesis.
A generic deformable 3D face model is built from the 3D scans of one hundred individuals. This generic 3D face model allows modelling the variation in the face-shape of different people. It can be fitted to the reconstructed 3D face-shape of an individual from a video sequence provided that the deviation from the real 3D face is less than certain thresholds. The methodology for reconstructing the 3D face-shape from a video sequence is provided. A structure-from-motion technique is used to estimate the poses of images in a video sequence. The estimated poses are optimized further by applying a bundle adjustment technique. Finally the dense correspondences between the images are estimated by employing Huber-L1 optical flow algorithm. With the optimized poseparameters and dense correspondences between the images, dense 3D face-shapes are reconstructed from a video sequence.
The application is developed to reconstruct the 3D face-shape in nearly uncontrolled environment where challenges exists due to varying lightning conditions, the non steady camera movement while capturing a video sequence and other factors using most promising state-of-the-art algorithm. Given the nearly uncontrolled conditions we were able to estimate a point cloud representing the face shape, but the results are not as accurate as expected. The related state-of-the-art algorithms used here have not been considered for such challenging conditions. The factors affecting the depth estimation in face region are analyzed and possible improvements to enhance the 3D face-shape reconstruction are presented within in this thesis.