Image-Driven Simplification and Optimization of Polygonal Models
Model simplification is often used in computer graphics to reduce the complexity of large polygonal models. For many graphics applications, the goal of this process is to produce a coarser model that is visually similar to the original, allowing the complex original model to be replaced with little or no loss in perceived quality. Existing simplification techniques, however, rely on geometry-based heuristics as an indirect indicator of visual similarity. We propose a novel image-driven simplification method that uses rendered images of a model and an image metric as a means of measuring visual similarity. Using this approach, changes in appearance due to shape, normals, color, texture, and visibility can be accounted for in a more direct manner, resulting in simplified models of high visual quality. Like most contemporary simplification methods, our method relies on edge collapse to coarsen the model. A sequence of such operations is performed in a greedy order, as determined by the image metric, such that the collapsed edge is the one that has the smallest visual impact on the image difference. For efficiency, the vertices introduced during simplification are positioned using geometric heuristics. This greedy, heuristical approach is not guaranteed to lead to the best possible model, however. In order to achieve maximum visual quality, the geometry and connectivity of the simplified model must be chosen so that each ve11ex and its neighbors are in an optimal configuration with respect to each other. We propose an image-driven mesh optimization technique that strives to achieve this goal, and which improves the visual quality of an already simplified mesh of fixed complexity. This method performs continuous optimization of vertex positions and surface attributes, while making local discrete connectivity changes to the mesh where appropriate. We show that this global optimization method consistently results in an improvement in the visual quality of simplified models, regardless of the simplification method used to construct them.