Institute for Data Analysis and Visualization

Image-Based Rendering is an exciting new field, which lies in between Computer Graphics and Computer Vision. We believe that the more we use the knowledge from Computer Vision in our graphics rendering algorithms, the better our final rendered images will be. This dissertation presents a framework to identify what information from computer vision is relevant for rendering and how to render it. Instead of only using the depth information per pixel, we compute what we call a depth uncertainty region around it. We show how to compute this region from an existing 3-D recovering algorithm called range-space search. We also present a new idea that further improve the uncertainty estimation, making it tighter. It is based on the assumption that the estimated depth is close enough to the actual surface. Two approaches for rendering are presented. The 4-D approach is based on the light field ray space parameterization. The 3-D approach is based on an existing technique of splatting 3-D Gaussian kernels. We show that both techniques achieve good results, but the 3-D approach is faster and it produces sharper images in most cases.