Towards 3D Facial Reconstruction from Uncalibrated CCTV Footage

Facial comparison in 2D is an accepted method in law enforcement and forensic investigation, but pose variations, varying light conditions and low resolution video data can reduce the evidential value of the comparison. Some of these problems might be solved by comparing 3D face models: a face model derived from CCTV camera footage and a reference face model acquired from a suspect. In our case we will assume uncalibrated CCTV footage, because the original camera setup may be destroyed or replaced after the incident, so precise camera information is no longer available. In contrast to other statistical methods, like Morphable Models, we would like to use no additional statistical information at all. Our method is based on a projective reconstruction of landmarks on the face and an auto-calibration step to obtain a 3D face model in a Euclidean space. In our experiment the effect of the number of frames and noise on the landmarks is explored for 3D face reconstruction based on landmarks. An estimation of the 3D face shape can already be obtained using 25 points in 30 frames.