Texture Generation from a Large Set of Registered Images Using Markov Random Fields

The process of texturing a 3D model from registered images is a well-known problem in 3D modeling and remote sensing. Recently large sets of registered images are common due to advances in Structure-from-Motion systems. We present a method for automatically texturing a 3D mesh from a large set of registered images. Occluding objects in the images are widely eliminated in the resulting texture by choosing one optimal source image for each triangle. Visible seams are avoided by considering neighboring triangles with a Markov Random Field approach. This allows an automatic generation of a texture while avoiding blurred textures that occur in image blending approaches. Manual intervention reduces to a fine tuning of factor weights for the image labeling process.