source: trunk/documents/theses/dstn/conclusion.tex @ 12301

The \an system applies the framework of geometric hashing to the
astronomical problem of blind astrometric calibration of images.  This
can be seen as an instance of object recognition in which the
individual objects to be recognized---stars and galaxies---are almost
completely indistinctive at the level of precision in typical images.
The geometric relationships between the objects, however, can be used
to build very distinctive features.  The problem is made easier by the
fact that the stars are very distant, so the viewpoint is essentially
fixed, and while the stars do move, their motions are small enough
that their geometric relationship change very little.  The problem is
difficult largely for its sheer scale: typical images cover
one-millionth of the surface area of the sky or less, and errors of
various types mean that stars are lost and gained in both the image
and the reference catalog of stars.


Chapters \chapref{chap:techreport}, \chapref{chap:verify} and
\chapref{chap:kdtree} highlight the main aspects of the system: a
geometric feature-indexing method that is able to generate
hypothesized matches; a robust probabilistic scheme for testing these
hypotheses; and a data structure implementation that allows the whole
system to operate at a speed that is acceptable for it to be used as a
practical tool.


We launched the \an system as a web service and release all the code
quite early in its development, and while operating and maintaining it
took considerable effort, we learned a great deal about real-world
data, and the fact that real people were using it to do real work
motivated us to continue tuning and refining the system.  As more and
more people are using the system and adapting it to their own
purposes, we feel we are getting closer to our goal ``to help
organize, annotate and make searchable all the world's astronomical
information.''