Robust frontal view search using multi-camera constrained Isomap

Many 2D face processing algorithms can perform better using frontal or near frontal face. In this paper, we present a robust frontal view search method, and apply it to a scene where a rotating head is simultaneously captured by multiple cameras. Our method is based on manifold embedding using Isomap, with the assumption that with head pose being the only variable, the face images should lie in a smooth and low dimensional manifold. We constrain Isomap by (1) a spatio-temporal constraint which improves the neighborhood graph of Isomap to make the manifold retain smooth and low-dimensional under uncontrolled conditions; (2) a multi-camera constraint which reflects the relative pose of multiple simultaneous video streams to adjust the graph distance of manifold in iteration. Using the constrained Isomap, the frontal view is in the vertex of the parabola-shaped manifold in 2D embedding. The experiments show that the results have high reliability.