COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	Here are some notes to supplement Saturday's discussion.  I will not summarize
C00011 ENDMK
C⊗;
Here are some notes to supplement Saturday's discussion.  I will not summarize
what was said then.

	As noted in the annotated copy of the thesis that I will give you, I
think that the beginning is unsatisfactory.  Let me suggest something like this:

	This dissertation is concerned with finding the correspondences between
points in narrow and moderate angle stereo pairs of points using methods that
do not involve the recognition of objects in the pictures or even the regognition
of the boundaries of objects standing out against the background.  The methods used
involve matching of small regions by translating to maximise correlation, together
with a large number of heuristics to reduce the impractically long searches
that the most straightforward application of the method would entail.  One of the
methods involves the determination of a camera model and by matching a small number
of points and using it to reduce two-dimensional to one-dimensional searches.

	There are two justifications for exploring methods of picture correlation
that do not involve object recognition:

	First, the state of the art in object recognition is not very far advanced,
and the techniques so far used only recognize a few kinds of objects.

	Second, while the human vision system makes good use of object recognition
when recognizable objects are in the picture, it also has considerable ability
to match stereo pairs of unfamiliar material.  Not using edge recognition techniques
is less defensible, but there turned out to be plenty to do without it, and also
it is possible to go quite far without it.

	Third, while an ultimate system will undoubtedly make use of methods that
do and do not use object recognition, it is a reasonable research strategy at present
to look at the methods separately as they involve quite different kinds of
programming.  

	The programs written as part of this research have performed about as
well as might have been expected from theoretical considerations.  However,
they are far from satisfactory for many of the applications envisaged, because
they are too slow and because they miss correspondences between small or thin
or unfeatured regions.  However, we are far from exhausting the possibilities
of the non-recognition methods, and we propose extensions that will reduce the
time required for the matching and will match regions not presently matched.
Further development of the correlation methods appears worthwhile.

	The direction of the research has been influenced by considering certain
potential applications.  These include computer driving which involves a richer
visual environment than laboratory or industrial manipulation tasks and the
Mars rover problem which requires a program that can drive a vehicle safely
when human intervention is limited by a round trip signal time of many
minutes.


--------------------

	You may use any of the above prose you want to assuming you agree with it,
but this is my idea of the points that should be covered in the introduction.
The points about the results and prospects of the approach taken are given in
about the level of detail for an introduction, although a few concise sentences
inserted at particular places might make things clearer.  The same points about
where the line of investigation now stands should be made in more detail in the
concluding section.  Another appendix on how to work the programs might be worthwhile
if it is not too hairy.
The essence of the introduction and conclusion is that they must tell the reader
the status and prospects of this line of work, and the thesis should support
the conclusions drawn.  Of course, your opinion of this is probably not identical
to what I have stated, and what is written should express your opinion, but you
can't leave the matter vague.  If you have some ideas of what cannot be done
without going to edge recognition or object recognition, you should state them
as precisely as you can and support them to the extent that you can.

	In general, your writing style is good, although somewhat redundant.  There
is no need to pad out a thesis to any particular length.  (Mine was 23 pages counting
an appendix, and I have seen a six page thesis in mathematics.)  Often, you
could change a sentence to give much more precise information without lengthening
it much.

	I don't like the present introductory page at all.  It tells the reader
nothing he doesn't know already, and he may not agree that the importance of
computer science stems from the use of computers to send bills.  As you see from
my draft, I favor plunging right in, but maybe this is extreme.  Perhaps the
introduction should be appropriate for that physicist who was the chairman of
the orals committee.

	However, there should be a half a page of mention of previous work, and
you should get the facts straighter.  For many years computer vision research
was dominated by local operations and by pattern discrimination.  The first
description based work was Roberts' thesis, done at MIT Lincoln Laboratory
and not at Project MAC.  Your acknowledgments of previous work should be precise.
In particular, you should acknowledge Gennery by name and make more precise
what programs and ideas of Quam's you are using.  To the extent that your
work is based on my ideas or that of others in the lab you should say so.  One
detailed sentence per acknowledgment is probably sufficient.

	1. Your opinion about whether FFT has a place should be stated and
supported better.  This will probably satisfy Baskett.  Spend some thought
on determining precisely what your opinion is.

	2.