Landmark-based model-free 3D face shape reconstruction from video sequences

In forensic comparison of facial video data, often only the best quality frontal face frames are selected, and hence potentially useful video data is ignored. To improve 2D facial comparison for law enforcement and forensic investigation, we introduce a model-free 3D shape reconstruction algorithm based on 2D landmarks. The algorithm uses around 20 landmarks on the face and combines the structure information of multiple frames. Model based 3D reconstruction methods, such as Morphable Models, reconstruct a 3D face shape model that is strongly biased towards the average face. Therefore, we don't use statistical face shape models in our model-free approach. The 3D landmark reconstruction algorithm simultaneously estimates the shape, pose and position of the face, based only on the fact that all images in the sequence are recorded using a single calibrated camera. The algorithm iteratively updates the reconstruction by including new frames, while maintaining the consistency of the reconstruction. We demonstrate the convergence properties of the method reflected in the 2D reprojection error and the 3D error with respect to a ground truth model. We show that the quality of the reconstruction depends on the level of noise in the landmarks. In follow-up experiments we show that our method is able to reconstruct the 3D structure of a face, using a styrofoam head and real video data. The results of the real face data show the same behavior as the results of the simulated data, which indicates that our method is capable of reconstructing real facial structures, depending on the noise of the landmarks.