Copy Link
Add to Bookmark
Report
VISION-LIST Digest 1988 07 12
Vision-List Digest Tue Jul 12 17:30:20 PDT 1988
- Send submissions to Vision-List@ADS.COM
- Send requests for list membership to Vision-List-Request@ADS.COM
Today's Topics:
Model-Based Vision (long)
CALL FOR PAPERS: Israel National Conference on AI
Looking for an article by Edwin Land
Re: Looking for an article by Edwin Land
----------------------------------------
Date: 1 Jul 88 20:39:01 GMT
From: jadwa@henry.cs.reading.ac.uk (James Anderson)
Subject: Model-Based Vision
Organization: Comp. Sci. Dept., Reading Univ., UK.
It seems to me that, once again, computer vision is in the grip
of the bottom-up, image processing camp. As an antidote to this I
propose the "strong model-based vision thesis" which sets out a
paradigm for computer vision and gives it a motivation. I would
like to invite comment on both of these aspects.
This is a long article, so if you are not interested stop
reading now!
____
In model-based vision an object is recognised by matching a
geometrical model of the object to an image. This is as far as
most people take it, but the models could, potentially, support
many more tasks, such as planning robot motions, learning new
models, and mediating between low-level image descriptions and
high-level, cognitive knowledge. These notions are expanded
below, but first, a definition.
The strong model-based vision thesis is that:
1) "Geometrical models of objects, together with lighting
models, support the automatic generation of all image
features needed to match a model to an image."
This requires that we develop a computational theory of image
formation, implement it in a vision system and make it available
to a meta-program which generates the image features for any
particular model. This is a lot to ask, but it does provide a
spur to developing formal theories of vision.
It also has a practical aspect. If a vision system can generate
the features it needs to find in an image to confirm a model
match then the program can be extended to deal with new objects
and new lighting conditions. We can also test the generality of
the system by manipulating these parameters.
A general program might also be optimised for particular,
restricted sets of models and or lighting conditions. Given a
well developed theory of model-based vision many such down-grade
paths can be found.
Such general, and therefore portable vision programs could be
exchanged between AI labs giving researchers a better opportunity
to judge their performance. This might sound like it is a
long-way off right now, but the image processing community
already publish packages of their software. The model-based
vision community (few and thinly spread) could donate model-based
feature evaluation software and thereby obtain a toe-hold in
published software. In the longer term, a well designed modelling
package might support common research, but if it is to do this we
must identify the features required of a modelling system to
support computer vision (hence this discussion).
Now, another definition.
2) "Visual knowledge is knowledge which can be brought into a
two-way spatial mapping with an image."
By definition, vision is the process of achieving a mapping from
an image to a knowledge representation. This, bottom-up, mapping
is sufficient for recognition, which is why it has received more
attention than the second kind of mapping.
A top-down mapping from knowledge to an image is not a necessary
property of vision in general. As I have just said, it is not
needed to recognise objects. But matching a model to an image
involves predicting the spatial location of features, therefore,
the top-down mapping is a necessary property of model-based
vision.
Also, we can explain human abilities, like pointing to an object
in a scene, most parsimoniously by supposing that the viewer sets
themselves the task of pointing to the object, say, a horse and
maps downward from the cognitive label, "horse in the scene" to
its actual location.
It is possible that semantic models of objects could be matched
to image features that carry no spatial description, but the
features would have to have tremendous discriminatory power. I
think it more reasonable to suppose that spatial location is an
important feature in most "natural" recognition tasks and,
therefore, matching knowledge to spatial descriptions is an
essential ability.
We have just seen that visual knowledge, as defined, is a
necessary part of model-based vision. I suggest that:
3) "Models mediate the mapping between images and visual
knowledge".
In addition to its tautologious necessity, I suggest that there
is a practical use of (3). Suppose that a model-based vision
system has analysed an image and we wish to check its
interpretation. We could ask for an interpretation of
everything, but are more likely to want to ask questions
about particular regions in the image. So we point to them with
a mouse, light pen, or some such, then the system maps from the
image location on to its models and tells us the model-based
interpretation of that location (it says what models are there).
We might also ask linguistic questions, such as, "Where is the
horse" or "Show me the image region that you interpret as the
edge defining the near side, back leg of the horse". In both
cases, models mediate the system's response.
As a follow on to (3) I suggest:
4) "Visual knowledge is related to all other kinds of
knowledge".
By this I mean that knowledge of the visual form of objects is
related to cognitive knowledge of their function, uses and even
emotional reactions to the objects.
Let me summarise the argument.
Tenets, (1-4) state a theory of vision, the strong model-based
vision thesis which promotes the development of formal,
automatically computable, theories for all aspects of model-based
vision.
Practical, paradigmatic, consequences follow.
A) That model-based vision programs predict image features
automatically and need only a new model and knowledge of
the lighting conditions to recognise new objects. This
makes them highly portable and testable.
B) That model-based vision programs can be interrogated by
pointing at an image or by asking questions. This makes
them highly testable and encourages the creation of good
software development environments.
C) That there is a downgrade path from a well formed
model-based vision system to a particular applications
system. This makes it even more worth while to implement a
modelling system specifically for vision.
James
(JANET) James.Anderson@reading.ac.uk
Computer graphics isn't the same as computer vision.
Let me show you what I mean ...
----------------------------------------
Date: Wed, 6 Jul 88 21:21:57 JDT
From: Shmuel Peleg <peleg%humus.Huji.AC.IL@CUNYVM.CUNY.EDU>
Subject: CALL FOR PAPERS: Israel National Conference on AI
CALL FOR PAPERS
Israel National Conference on AI
Tel-Aviv, Dec 27-29, 1988
Papers are solicited for the 5th National Conference on AI. The
conference is being held, this year, in conjunction with the Vision,
Image Processing and Pattern Recognition SIG of ILLA, the Israeli
Association for Information Processing.
The conference will take place on December 27-29 1988, and there will
be a special track for Vision and Image Processing related issues.
Topics of interest for the vision program include:
- Robot and machine vision
- Image processing and understanding
- Graphics
- Modeling of human perception
- Pattern recognition and analysis
- Parallel algorithms for IP applications
- Industrial and Biomedical application
For further information please contact Dr. Hezy Yeshurun, Dept of
Computer Science, Tel Aviv University. Phone : 03-5413553
E-mail: hezy%math.tau.ac.il@relay.cs.net, a13@taunos.bitnet
----------------------------------------
Date: 6 Jul 88 09:35:34 GMT
From: mcvax!prlb2!ronse@uunet.UU.NET (Christian Ronse)
Subject: Looking for an article by Edwin Land
Keywords: Scientific American, 1959.
Organization: Philips Research Laboratory, Brussels
Could someone give me the exact reference (Volume, number, pages, month) of
a paper by Edwin Land which appeared presumably in 1959 in Scientific American?
It is entitled ``Experiments in Colour Vision''.
Thanks.
Christian Ronse maldoror@prlb2.UUCP
{uunet|philabs|mcvax|cernvax|...}!prlb2!{maldoror|ronse}
----------------------------------------
Date: 8 Jul 88 13:14:40 GMT
From: mcvax!prlb2!quisquat@uunet.UU.NET (Jac. Quisquater)
Subject: Re: Looking for an article by Edwin Land
Keywords: Scientific American, 1959.
In article <459@prlb2.UUCP> ronse@prlb2.UUCP (Christian Ronse) writes:
>
>Could someone give me the exact reference (Volume, number, pages, month) of
>a paper by Edwin Land which appeared presumably in 1959 in Scientific American?
>It is entitled ``Experiments in Colour Vision''.
>
>Thanks.
>
>Christian Ronse maldoror@prlb2.UUCP
>{uunet|philabs|mcvax|cernvax|...}!prlb2!{maldoror|ronse}
Volume 200,
Number 5,
May, 1959
pp. 84-99
Fascinating book
...
available
...
in my office!
Abstract (my answer is too short for this @#$%^&-= postnews):
The eye [...] can perceive full color in images which, according to
classical theories, should be monochromatic.
------------------------------
End of VISION-LIST
********************
