Let's suppose I would like to to recognize automatically images like:

1) Someone is smoking from somene who is not smoking
2) A cigarette or a cigarette box from some other object
4) A weather forecast from any other image
5) A male face from a female face.

What softwares are available? What toolboxes? What libraries?

Please avoid any research paper or code, that aim at demonstrating the
feasibility of a new technique in a particular context and with specific
data.

List only practically usable packages or pieces of code, included well
structured and documented C, Java and Matlab codes, with an easy
possibility to get trials, demos and/or limited cost versions, or some
other proof that at least one of the points above is practically
feasible by installing proper software and integrating it somehow with
other image acquisition and communication softwares.

Please add any available reliable judgement or review on the listed softwares.

---

Request for Question Clarification by endo-ga on 29 Sep 2004 04:46 PDT

Hi,

Do you want to use Windows or do you have no preference?

How practical do you want the examples or the code?

Because you will still have to code quite a lot and program the
algorithms even if you do know what tools can be used.

Thanks.
endo

---

Clarification of Question by edddh-ga on 29 Sep 2004 08:27 PDT

1) Any OS is acceptable, but Windows can make the initial study and
experimentation easier.

2) I am expecting to have real software packages, not theories or
subroutines. One must be able to easily and practically reproduce
relevant examples, i.e., install the required software, perform a
learning with an existing or new set of example images, and categorize
input images obtaining a satisfying degree of success.

---

Request for Question Clarification by endo-ga on 30 Sep 2004 03:59 PDT

Hi,

Thank you for your clarification, I doubt that such software exists
'out of the box'.

The problems you describe are not trivial, and there are researchers
that have probably created algorithms to perform those operations.
Therefore if you're willing to implement an algorithm using image
manipulation software then you can achieve your goals. Otherwise you
might need to wait a while for someone to write commercial versions.

Thanks.
endo

---

Clarification of Question by edddh-ga on 30 Sep 2004 04:38 PDT

Thanks a lot for your request of further clarification.

If you can somehow guarantee that this software does not exist 'out of
the box', you could certainly list and describe the libraries and
modules that could be used to solve at least one of the listed
problems, together with a list of research or other institutes that
know how to achieve the final result adding the missing necessary code
to those libraries.

The concern expressed in the text of my question comes from the
thought that you need to have some kind of proof (from trials or
demos) to be able to reach the practical result in a definite number
of steps, but not necessarily in short time and/or excluding coding
and integrations. I am referring to the following note:

"List only practically usable packages or pieces of code, included well
structured and documented C, Java and Matlab codes, with an easy
possibility to get trials, demos and/or limited cost versions, or some
other proof that at least one of the points above is practically
feasible by installing proper software and integrating it somehow with
other image acquisition and communication softwares."

Thanks again, and I hope you can write an answer with my last clarification.

As a basic approach, you could take FFTs of the images and compare
them statistically... it is doubtful that such an approach would yeild
very accurate results, but it may be accurate enough for your needs.

More likely, though, you'll want to use a neural net package and then
spend a lot of time training it on the specific distinctions you want
it to be able to make.  I would consider 60% accuracy to be fairly
strong results for the problems you describe (best case)... the human
eye is a miraculous thing, and the brain processes so much information
in order to come up with visual distinctions and pattern recognitions
that we take for granted... it's mind boggling.

Have a look at MediaBin from Interwoven.com.  It allows you to say
"show me all other images similar to this one", etc.  A pricey beast,
though.

Hi

I am quite sure that there is no such software available at the moment.
I am researcher in this field and software like MediaBin won't help
much to your problems.

What you definitively need is a solution specially tailored to your
problem and this is only possible if you have lot of training data.

Although the problem you are addressing is of high interest in the
computer vision community at the moment (e.g.
http://www.pascal-network.org/challenges/VOC/).

hello,

during my phd thesis, i developped a method for image
classification/categorization. it was evaluated on various
image datasets (faces, digits, objects, landscapes, themes,
buildings, ...). 

in collaboration with PEPITe, we developped a Java software
called "PiXiT" and based on that method:
http://www.montefiore.ulg.ac.be/~maree/pixit.html

here are our publications if you want more scientific information:
http://www.montefiore.ulg.ac.be/~maree/#publications

please feel free to ask me more details.

raphaël.