Textured 3-D face reconstruction with two images in different views

In this paper, we present a new approach for textured 3-D face reconstruction based on two images in different views, which requires minimal user interaction by selecting eight pairs of feature points corresponding to each of the two images. These feature points have three functions: computing the fundamental matrix with the two face images based on epipolar geometry, estimating the poses of the images with respect to the pose of the frontal view, and determining the positions of the features for a rough adjustment of a generic 3-D face model. After the epipoles have been obtained, with the eight pairs of points as seeds, more matching points across the two images are detected under the constraints of epipolar lines. Then, the 3-D coordinates of the matching points can be computed with an SVD algorithm. After that, the generic 3-D model is adopted to optimize the constructed face model by eliminating the noise. Finally, the image closer to the frontal view is chosen as the texture map of our 3-D face model. Compared with some other previous works, there are fewer constraints to the poses of the images in our system and only one generic 3-D face model is needed. Experimental results demonstrate an encouraging performance of our method.