Lost but Found? Harnessing the Internet for Photometric Completion

While it is important to digitize heritage sites 'as is', building 3D models of damaged archaeological structures can be visually unpleasant due to the presence of large missing regions. This work addresses intensity filling-in, or intensity inpainting, of such large damaged regions post geometric reconstruction. Assuming a Lambertian image formation model, we first establish that patches corresponding to arbitrarily oriented planar regions found in internet images of several archaeological structures with possibly different albedo and observed under varied and uncontrolled illumination lie in a low-dimensional subspace. These are then used for inpainting by modeling the missing region as gross pixel corruptions. The performance of the proposed method along with comparisons are shown on synthetic as well as real data.