Global Multi-modal 2D/3D Registration via Local Descriptors Learning