Message-ID: <223356Z31081994@anon.penet.fi>
Path: cantaloupe.srv.cs.cmu.edu!das-news.harvard.edu!news2.near.net!MathWorks.Com!udel!gatech!howland.reston.ans.net!EU.net!news.eunet.fi!anon.penet.fi
Newsgroups: comp.ai
From: an62757@anon.penet.fi (Marr born again)
X-Anonymously-To: comp.ai
Organization: Anonymous contact service
Reply-To: an62757@anon.penet.fi
Date: Wed, 31 Aug 1994 22:27:51 UTC
Subject: The best approach for object recognition
Lines: 38

Over the years, people have proposed several different approaches for 
object recognition. Some that come to my mind right now:
(1) Geon based approach (popularised by Biederman)
(2) Extracting a hierarchical description of objects and then matching
as described by Marr & Nishihara [As it appears to me, the main differences 
between this approach and the Geon based approach are  that  (a) Geon based
approach is much more qualitative in nature while the former incorporates
significant amount of quantitative information, (b) Geon based approach does
not involve computing depth maps (c) Geon based descriptions are not 
hierarchical. In the examples I have seen, geons haven't been subdivided into
``sub-geons''.].
(3) The alignment approach [Popularised by Lowe, Ullman].
(4) The interpolation approach [Popularised by Poggio, Edelman & group],
[As I see it, this is quite similar to the alignment approach except that
(a)One does not try to match the image with just one model(view) but 
with a bunch of them, (b) One does not explicitly compute the transformation
parameters.]

Now, some questions:
(A) Is there any other major approach that I am missing? [Forget about 
approaches such as (1) Matching Fourier transforms (2) Matching moments.
It seems pretty convincing that these approaches will work only in very
specific circumstances].
(B) What are the main advantages and disadvantages of these approaches? 
What do you (an individual netter) consider to be the most promising approach?
Why?

I suggest that we discuss these issues at length; it will prove very useful.
While there have  appeared quite a few survey papers, most of them just give
the feeling of being a collection of independent paragraphs, with one 
paragraph allocated to each ``major'' paper/project. I think it would be nice
to get a critical overall view of the major approaches towards object 
recognition.  
-------------------------------------------------------------------------
To find out more about the anon service, send mail to help@anon.penet.fi.
Due to the double-blind, any mail replies to this message will be anonymized,
and an anonymous id will be allocated automatically. You have been warned.
Please report any problems, inappropriate use etc. to admin@anon.penet.fi.
