Constrained Up-Scaling for Direct and Global Image Components

Abstract

The separation of direct and global illumination components is interesting for many applications in Computer Graphics and Computer Vision, such as BRDF estimation or material classification. However, for full-resolution images, a large number of coded images have to be acquired. For many interactive applications, such as the acquisition of dynamic scenes or video capturing, this is not feasible. In this paper, a new constrained up-scaling technique for separated direct and global illumination images is proposed which requires two to three coded input images, only. Our approach imposes the boundary condition that the sum of the direct and global components equals the fully illuminated image. We work in a predictive-corrective manner where we first use a single-image up-scaling method in order to predict the higher resolution images. Afterwards, the missing higher frequencies are determined using a fully illuminated image. As the distribution of the higher frequencies differs among the various frequency bands, we apply our approach in an iterative way for small up-scaling steps distributing the missing information by minimizing the overall frequencies. We evaluate the up-scaling scheme and demonstrate the improvement compared to single-image approached. As our method aims at minimizing the structured light patterns needed for acquisition, we additionally discuss the performance of existing pattern sets in terms of applicability for dynamic scenes.