Unfortunately the implementation has not yet reached a point where reliable run times and similar statistics can be reported. Instead some qualitative observations are presented. Some images (figure 7) further document the results.
The deviation spaces (section 4.1) succeed in adding detail near the silhouettes as reported in . While this works well for regions with low curvature, the deviation spaces degenerate to spheres after few simplification levels for rough regions. This is not surprising, but an algorithm that delays the degeneration to higher levels as far as possible would be desirable.
Geomorphing as described in section 4.2 produces very attractive results. The only drawback of the static approach seems to be the presence of many nearly degenerate triangles under certain circumstances. However, this is only visible in wireframe mode because texture coordinates and normal vectors are interpolated together with the vertex locations.
Parallel preprocessing was tested on a Dual-Pentium-II board. A rough estimation of run times revealed that about 8% of the simplification code run in critical sections (i.e. only one thread at a time). That results in a lower bound of 54% of execution time compared to a uniprocessor machine. In experiments the execution time could be reduced to 64%. While this result is not too bad for a two-processor machine, some improvements might increase parallelism. At the moment the priority queue seems to be the bottleneck when using the concurrency scheme of section 4.3.
No detailed statistics are available about progressive transmission (section 4.4) by now. First experiments indicate that transmission takes place in reasonable times even at a low bandwidth. An integrated utility allows the bandwidth to be explicitly limited so that the effect of using modem lines and other typical transmission media can be explored.