It's possible to derive some conclusions about the accuracy from the assumptions stated on the main research project page.
The accuracy of the system is only as good as the data captured from reality. For example it's not possible to detect fully occluded objects with a normal camera. Also a wrong focal distance or bad lighting will hinder good tracking.
The accuracy of the system is only as good as the data generated by the computer. For example it's just a rough assumption of real light conditions to use a single one point light and standard OpenGL based rendering.
Note: The simulation must include the whole evolutionary story of the image. If an camera captures images with a radial distorition, the renderer should do the same.
The accuracy of the system is only as good as the mechanism which compares generated and captured datasets. Completly identical data must result in 100% similarity. That's easy. The tricky part is to get a good statement about almost mathching and partial matching data. See the selection page for a discussion.