In this talk, I will show how Machine Learning can help solving the well-known Shape-from-shading problem. This problem is usually solved using either methods of resolution of the PDEs or optimization techniques. Our novel approach requires to use a collection of spheres with varying radii. More precisely, the images of these spheres are cut into small patches. Afterwards, the image which constitutes the datum of the problem is cut into patches of the same size. Each such patch is associated to the "nearest" patch in the learning set. Its normal is then straightforwardly estimated. Some examples will be presented, which show that this alternative approach is worth of interest.