Hi everyone,

I've had an idea as to a method of using an automatic algorithm to build
images from photos of a complex object, and was interested to see what
others thought of it...

The basic idea is using an 'evolving' system to add multiple objects to a
scene, working like most adaptive algorithms

1. Start off with a first guess approximation
2. Add random noise in the form of small objects/blobs in a different manner
to say 1000 test scenes.
3. Test each scene for similarity to the photos
4. Take the highest scoring models, and 'breed them' merging commands, then
repeat from step 2.

Run for several tens of thousands of iterations (anyone got a spare super
computer?) The lighting will have to be altered as well as the object
patterns...

What do y'all think? Any suggestions? Anyone with any experience in this
kind of coding?

Just a thought,

Regards,

Luke

Post a reply to this message

On Tue, 29 Jan 2002 19:57:03 -0000, "Luke Church"
<luk### [at] hotmailcom> wrote:

>What do y'all think? Any suggestions? Anyone with any experience in this
>kind of coding?

Given almost infinite time, what you'll probably end up with is a
large amount of small objects, each taking a pixel (or several), which
will only make sense when looking from a specific point in 3D space.
You can do that with a single iteration so why bother? :)


Peter Popov ICQ : 15002700
Personal e-mail : pet### [at] vipbg
TAG      e-mail : pet### [at] tagpovrayorg

Post a reply to this message

> Given almost infinite time, what you'll probably end up with is a
> large amount of small objects, each taking a pixel (or several), which
> will only make sense when looking from a specific point in 3D space.
> You can do that with a single iteration so why bother? :)

Certainly true if only one perspective is used. Sorry, I wasn't clear... I
was considering using at least 3 photos of the same object under similar
lighting conditions taken ideally at equal intervals around the object and
comparing all of them to the model...

The original idea for this was effectively a poor mans digitiser using an
algorithm to make up for not being able to afford to build proper sensing
equipment, with the hope of arriving at a more 'real looking' model of a
complex object that would otherwise be practical...

Thanks for the input, :)

Luke

Post a reply to this message

On Wed, 30 Jan 2002 00:03:55 -0000, Luke Church wrote:
>> Given almost infinite time, what you'll probably end up with is a
>> large amount of small objects, each taking a pixel (or several), which
>> will only make sense when looking from a specific point in 3D space.
>> You can do that with a single iteration so why bother? :)
> 
> Certainly true if only one perspective is used. Sorry, I wasn't clear... I
> was considering using at least 3 photos of the same object under similar
> lighting conditions taken ideally at equal intervals around the object and
> comparing all of them to the model...

You could still converge to disjoint clouds of particles, each of which only
appears in one of the three (or whatever) views of the scene.

-- 
plane{-z,-3normal{crackle scale.2#local a=5;#while(a)warp{repeat x flip x}rotate
z*60#local a=a-1;#end translate-9*x}pigment{rgb 1}}light_source{-9red 1rotate 60
*z}light_source{-9rgb y rotate-z*60}light_source{9-z*18rgb z}text{ttf"arial.ttf"
"RP".01,0translate-<.6,.4,.02>pigment{bozo}}light_source{-z*3rgb-.2}//Ron Parker

Post a reply to this message

In article <slr### [at] fwicom>,
 Ron Parker <ron### [at] povrayorg> wrote:

> You could still converge to disjoint clouds of particles, each of which only
> appears in one of the three (or whatever) views of the scene.

That might actually be quite interesting...sort of an abstract sculpture 
with an actual point. It might be worth making something that does 
this...not a POV script, but a separate program that does a bunch of 
renders with scanlining or some other technique to figure out the 
information and then outputs the data, either as particles or voxel 
data. It could save every 100th frame or so, and might produce results 
within a reasonable period of time...

-- 
Christopher James Huff <chr### [at] maccom>
POV-Ray TAG e-mail: chr### [at] tagpovrayorg
TAG web site: http://tag.povray.org/

Post a reply to this message

A lot of work has already been done with generating 3d geometry from 
multiple 2d images using techniques such as structured light.
http://www.prip.tuwien.ac.at/Research/3DVision/struct.html


Luke Church wrote:

>> Given almost infinite time, what you'll probably end up with is a
>> large amount of small objects, each taking a pixel (or several), which
>> will only make sense when looking from a specific point in 3D space.
>> You can do that with a single iteration so why bother? :)
> 
> Certainly true if only one perspective is used. Sorry, I wasn't clear... I
> was considering using at least 3 photos of the same object under similar
> lighting conditions taken ideally at equal intervals around the object and
> comparing all of them to the model...
> 
> The original idea for this was effectively a poor mans digitiser using an
> algorithm to make up for not being able to afford to build proper sensing
> equipment, with the hope of arriving at a more 'real looking' model of a
> complex object that would otherwise be practical...
> 
> Thanks for the input, :)
> 
> Luke

Post a reply to this message

On Wed, 30 Jan 2002 00:03:55 -0000, "Luke Church"
<luk### [at] hotmailcom> wrote:

>Certainly true if only one perspective is used. Sorry, I wasn't clear... I
>was considering using at least 3 photos of the same object under similar
>lighting conditions taken ideally at equal intervals around the object and
>comparing all of them to the model...

Ah, image reconstruction... there's a lot of info on that on the net.
For now, given almost infinite time, stick with this scenario:

1. Make a prism or cone with its vertex at location and its base
extending about as much beyond look_at. Have it pass through a pixel
of the image and be of approximately the same visual size.

2. Repeat for all pixels

3. Union all those cones

4. Repeat 1-3 for all viewpoints

5. Intersect all those unions

6. Grow old waiting :)


Peter Popov ICQ : 15002700
Personal e-mail : pet### [at] vipbg
TAG      e-mail : pet### [at] tagpovrayorg

Post a reply to this message

Ron Parker writes:
> You could still converge to disjoint clouds of particles, each of which only
> appears in one of the three (or whatever) views of the scene.

Applying a penalty to the "fitness" of the generated images which
increases as the number of objects increases might avoid that.  Large
render times could also be penalised.  Of course, that may increase
the number of iterations required.

Rohan

Post a reply to this message

Ron Parker wrote:
> You could still converge to disjoint clouds of particles, each of which only
> appears in one of the three (or whatever) views of the scene.

But you can solve that just by using more reference photos. Might look
kewl anyhow!

-- 
signature{
  "Grey Knight" contact{ email "gre### [at] yahoocom" }
  site_of_week{ url "http://digilander.iol.it/jrgpov" }
}

Post a reply to this message

Not so weird, use a genetic algorithm and remember to write a fitness 
function which in some way ranks better less complex solutions (otherwise 
you really end with a sphere for each voxel...).

By the way, it is a really interesting project!

Bye!!!
        Alessandro Coppo
        a.c### [at] REMOVEiolit

Post a reply to this message