Robust identification and matching of fiducial points for the reconstruction of 3D human faces from raw video sequences

We describe an algorithm for the automatic features detection in 2D color images of either frontal or rotated human faces. Such features allow to initialize robustly a bundle-adjustment in order to fit a generic 3D face model to the images. The algorithm first identifies the sub-images containing each feature (eyes, nose and lips), afterwards, it processes them separately to extract fiducial points. The features are looked for in down-sampled images, the fiducial points are identified in the high-resolution ones. The method uses both color and shape information and does not require any manual setting or operator intervention. It has been tested on a database of 130 color images.