This paper describes a model-assisted system for reconstruction of 3D faces from a single consumer quality camera using a structure from motion approach. Typical multi-view stereo approaches use the motion of a sparse set of features to compute camera pose followed by a dense matching step to compute the final object structure. Accurate pose estimation depends upon precise identification and matching of feature points between images, but due to lack of texture on large areas of the face, matching is prone to errors. To deal with outliers in both the sparse and dense matching stages, previous work either relies on a strong prior model for face geometry or imposes restrictions on the camera motion. Strong prior models result in a serious compromise in final reconstruction quality and typically bear a signature resemblance to a generic or mean face. Model-based techniques, while giving the appearance of face detail, in fact carry this detail over from the model prior. Face features such as beards, moles, and other characteristic geometry are lost. Motion restrictions such as allowing only pure rotation are nearly impossible to satisfy by the end user, especially with a handheld camera. We significantly improve the robustness and flexibility of existing monocular face reconstruction techniques by introducing a deformable generic face model only at the pose estimation, face segmentation, and preprocessing stages. To preserve data fidelity in the final reconstruction, this generic model is discarded completely and dense matching outliers are removed using tensor voting: a purely data-driven technique. Results are shown from a complete end to end system.
