Preprocessing

Given a set of visualization parameters such as compositing technique in use, etc., the goal of the preprocessing step is to classify voxels of the volume into voxels which possibly contribute to an image and voxels which do not contribute to an image. The classification criteria depend on the chosen opacity transfer function, the desired compositing method, and the degree of freedom which should be provided for further manipulation of the transfer function. Generally speaking, the more voxels are classified as irrelevant, the fewer data has to be processed during projection, and the faster the rendering gets.

If an iso-surface should be rendered (characterized by a sharp transition between transparent and opaque voxels at the iso-value), only voxels close to the surface of the iso-value transition are relevant. Voxels with lower values are entirely transparent, voxels with higher values which are located inside the surface do not contribute, as they are occluded by opaque voxels at the surface. Of course, this relevance condition is bound to a specific iso-value - specifying a new iso-value to view another surface makes reclassification of the voxels necessary.

A higher flexibility with respect to transfer function tuning is obtained by extracting all voxels of an object. Volumetric data sets often contain objects of interest which are surrounded by irrelevant data, for example, body parts surrounded by air in medical data sets (figure 3.1a), or attractors surrounded by ``empty'' regions of phase space in the case of dynamical system data (figure 3.1b).

Figure 3.1: Sections through volumetric data sets. A significant amount of voxels does not belong to any object of interest.

a) CT scan of a human hand	b) section through chaotic attractor

The viewing direction can also be included into the classification function. By subdividing the range of all possible viewing directions into several clusters, and performing classification for each cluster separately, several sets of relevant voxels are obtained. Each set is usually significantly smaller than the set of relevant voxels without considering viewing directions. As the criteria are highly dependent on the actual target visualization scenario, they will be presented in more detail together with the description of the rendering methods in chapter 4. The time required for classification is as variable as are the criteria. Simple conditions, like the identification of voxels which belong to a surface, require typically less than a second for an entire data set. Complex conditions, like applied for maximum intensity projection, may take several minutes of preprocessing.

mailto:mroz@cg.tuwien.ac.at.