Image-based rendering replaces the difficult physical simulation problem of conventional computer graphics with a difficult reconstruction problem. In this paper, we introduce efficient methods for eliminating common reconstruction artifacts in 3D image warping. These methods are suitable for hardware implementation. We study the reconstruction problem in the context of the 3D warp's application to post-rendering warping. This application provides us with perfectly controlled reference images and allows us to compare generated images against a gold standard provided by conventional rendering.
We characterize the reconstruction problem as one of reconstructing 2D manifolds (surfaces) in 3-space, followed by a projection, resampling, and compositing of these surfaces. We present two reconstruction and resampling algorithms. The first algorithm is a general one. The second algorithm is an efficient micro-polygon technique that uses flat-shaded, axis-aligned rectangles and super-sampled anti-aliasing to avoid explicit interpolation between re-projected samples. We also present an algorithm, based on epipolar geometry, to estimate the color of any unknown areas of the warped image. This algorithm's work is strictly bounded by the resolution of the warped output image. Finally, we discuss the effectiveness of our post-rendering warping system.