Answering Mike Walter's question:  A database of scanned titles was
always the intention of the IAMMP web site, http://www.iammp.org/ 
created as part of the rollscanners project by Warren Trachtman.
Scanner-owners would deposit their scans in it, and the database would
be populated using the metadata embedded as part of the processing.
The index is viewable by all, although general access to the contents
is restricted by copyright considerations to pre-1923 material only.

While the database hasn't been used as much as intended, it's not done
badly and as of this evening there are just over 9500 rolls listed.  It
has most of the categories Mike suggests, although not the material to
allow identification of variants of the same roll.

And as for the rollscanners-group scanned roll format, though -- it's a
Type 0 MIDI file, with special use of the MIDI data to represent piano
roll details.  There are two variants, the 'scanimage' that records the
results of processing the image of the roll, and 'punchmaster' that
stores the reconstructed original perforation pattern.  These differ
in their use of the timing within the file, the punchmaster file using
a fixed number of 'ticks' per punch row.  Both of these types include
all the perforation detail, including bridging, so aren't directly
playable.

The eroll format is a processed form of the roll for operating
pneumatic instruments: offset expression tracks are shifted, bridges
are stripped and it's re-timed to a standard step rate regardless of
the source, so it's not an archival form.  Warren's software creates
erolls in parallel with the 'scan' or 'punch' archival files.

What you'll find in the IAMMP public database is none of the above!
These files are simply playable MIDI files created from the roll. In
the case of 88-note rolls these will essentially be the same as erolls,
but for any rolls which use note tracks to operate expression devices
the playable MIDI file needs to store the resulting expression.

Julian Dyer