Evaluation of Memoryless Simplification

Error metric comparisons based on quality are typically dealt with by using visual image comparisons and model deviation measurement computations performed on the models after simplification. Performance measures of simplification techniques are measured based on computation time, regardless of the excess hardware resources used to improve these results. Comparisons of error metrics which are independent of algorithmic optimisations are not possible, since these optimisations make it impossible to use the same platform for different techniques. Due to the implementation specific nature of these simplification schemes, image comparison as a form of metric evaluation has only been performed visually on the simplified model produced. Although general error metrics provide good results during simplification, they may not successfully deal with surface attributes such as normals, texture coordinates and color values. Within the memoryless framework, new error metrics can quickly be devised and tested under the same conditions to determine which performs best with the model being simplified. This gives us the opportunity not only to evaluate various error metrics in terms of their performance, but also allows us to draw conclusions about how surface simplification is evaluated. It is difficult to define an effective measure of the deviation of a surface from the original model. Since models in three dimensions will be approximated on a screen in two dimensions, a two dimensional image-based comparison (from many viewing angles) would emulate how we would perceive error in the model. Unfortunately, an image based comparison measure has many parameters (the size of the image plays a large part in the magnitude of the error), and is difficult and slow to emulate on computers with no specialised rendering hardware. A number of techniques are available to assess model quality. Generally these modelbased techniques are considerably easier to compute than image based measures, as the are independent of the graphics hardware. We evaluate a number of models during simplification with both image-based and model-based measurements. Our results show that the rate of decrease in model volume corresponds closely with our image-based error measures.