I must admit I haven't tried the Stanford software, because I use
Warren Trachtman's programs to convert roll images to MIDI, and my own
software to simulate expression for Duo-Art, Welte and Triphonola.

The Stanford processing follows broadly similar steps to the earlier
work by Wayne Stahnke and Warren Trachtman and the others in the
'Rollscanners' project 15 or so years ago.  Basically, you start with
an optical scan creating an image of the roll at about 300 dpi, then
lay a tracker-bar map across the roll to determine what perforations
are present.

You then walk along the length of the roll creating a MIDI file that
is timed with one 'tick' per scan-line.  Stanford call this 'raw MIDI'.
There are two forms of this, one keeping any bridging and the other
removing it as a necessary step towards making a simulation.

Then, as Spencer Chase has described, come simulations of performances
from the raw MIDI.  These are always the hardest thing to do because
they need to model how the instruments respond to the roll coding,
which is mostly somewhat different to the tracker-bar descriptions.
So these need a thorough understanding of how instruments behave in
real life.

Basic simulations tend to omit the little bits about instruments that
are unsaid -- such as the slew-rate of their controllers (how fast
dynamics change as the coding changes).  No simulation can be better
than listening to an original instrument, they are useful tools to
preview scans or perhaps to help find faults in instruments.

Stanford's software doesn't contain a key step in roll-image
processing, which is to "recover the punch master" from the scan.
In this, knowledge about how rolls are made is added to the data
measured from the roll (a bit like OCR software overlays knowledge
about printing of text onto a scan of a document).  This removes
scanning errors (azimuth skew being perhaps the most significant)
and results in notes being precisely identified by their punch row:
a digital representation of the roll rather than an analog
approximation.

This processing is complex and time-consuming and few use it, another
reason simulations generally are rather weak.

Julian Dyer