Four-dimensional light-field modeling and rendering

Image-based models have recently become an alternative to geometry-based models for computer graphics. They can be formalized as specializations of a more general model, the light field. The light field represents everything visible from any point in 3D space. In computer graphics the light field is modeled as a function that varies over the 4D space of oriented lines.; Early models parameterize an oriented line by its intersection with two parallel planes, a parameterization that was inspired by holography. In computer graphics it introduces unnecessary biases that produce a rendering artifact called the disparity problem. We propose an alternative isotropic parameterization, the direction-and-point parameterization (DPP). We compare it to other parameterizations and determine whether they are view-independent, that is, invariant under rotations, translations and perspective projections. We show that no parameterization is view-independent, and that only the DPP introduces a single bias. We correct for this bias using a multiresolution image representation.; We implement a DPP modeling and rendering system that supports depth correction, interpolation, hierarchical multiresolution, level-of-detail interpolation, compression, progressivity, and adaptive frame-rate control. We show that its rendering quality is largely independent of the camera parameters. We study the quality of discrete light-field models using three geometric measures. Two quantify discretization errors in the positional and directional parameters of light field. The third measure quantifies pixelation artifacts. We solve three open problems: (i) how to optimally choose planes for two-plane and DPP models, (ii) where to position the discretization windows within those planes, and (iii) how to choose optimal window resolutions. For a given amount of storage, we show that DPP models give the best overall quality representation for 4D light-field modeling and rendering.; We demonstrate the application of 4D light-field models to holography. We generate a holographic stereogram based on both planar and isotropic representations. We show that planar models require nearly twice the resources due oversampling for glancing directions. Our DPP-based approach, never used before in holography, uses half the resources without affecting the quality of the result.