Multi-view Normal Estimation – Application to Slanted Plane-Sweeping

In this paper, we show how to estimate the normals of a 3D surface from a minimum of two views, assuming that the poses of a calibrated camera are perfectly known. For each pair of image points, the normal at the corresponding 3D point is expressed in function of the local gradients of the grey level, whatever the type of image formation (orthogonal or perspective projection). As an application, this allows us to fully estimate the inter-image homography, which not only depends on the relative pose between views, but also on the local orientation of the surface. Hence, the photo-consistency between patches from two images, which is the basis of the so-called “plane-sweeping” method, is improved. Experiments on synthetic and real data validate our approach.