Okay, not really talking about dedicated 3D input devices. What I mean to
ask is, has anyone thought of or perhaps even done modelling with
multiple-view camera setups?

What I have in mind is building a clay or plastic model then photographing
it using a digital camera. This setup would include the following:
1. the real-world model containing marked control points for surface
extrapolation
2. a camera, moving around the model in an arbitrary fashion, taking
pictures along the way (not necessarily in a pre-defined path)
3. a marker - a small model, such as a transparent cube, which would model
the origin and axes. This marker would serve as reference for the scale and
orientation for every photograph

Now, what I'd like the tools, with which to accomplish this task, to do is
to calculate the marked points on the model. This process would not be
fully automated, rather each image would first be reviewed by the user and
the user him/herself would mark the appropriate points on the photograph of
the model to define the 2D points which correspond to specific 3D points on
the model. Also, the user would need to mark the "origin cube" to define
reference points. After this stage of point-and-mark is done, the program
calculates the 3D points for each of the photographs and uses statistical
analysis to determine the most likely position of a point on the model.

I'm not looking for depth-of-field algorithms to produce dense 3D models;
just the points, so I can model, say, NURBS surfaces using those reference
points. I've looked at some sites offering some image rectification and
affine transformation with projection, but they're all built for dense 3D
reconstruction.

Here's one example:
http://www-sop.inria.fr/robotvis/personnel/qdelam/ArticleFG98html/Article.forhtml.html

It concerns moving a modelled hand to fit the depth-of-field produced by the
program.

Any information regarding these kinds of calculation? I guess I'm not alone
on this subject, who hasn't thought of simplifying modelling to, actually,
child's play? ;)

Thanks,

-Klemen

Post a reply to this message

Klemen wrote:
> Okay, not really talking about dedicated 3D input devices. What I
> mean to ask is, has anyone thought of or perhaps even done modelling
> with multiple-view camera setups?
>
> What I have in mind is building a clay or plastic model then
> photographing it using a digital camera. This setup would include the
> following:
> 1. the real-world model containing marked control points for surface
> extrapolation
> 2. a camera, moving around the model in an arbitrary fashion, taking
> pictures along the way (not necessarily in a pre-defined path)
> 3. a marker - a small model, such as a transparent cube, which would
> model the origin and axes. This marker would serve as reference for
> the scale and orientation for every photograph
>
> Now, what I'd like the tools, with which to accomplish this task, to
> do is to calculate the marked points on the model. This process would
> not be fully automated, rather each image would first be reviewed by
> the user and the user him/herself would mark the appropriate points
> on the photograph of the model to define the 2D points which
> correspond to specific 3D points on the model. Also, the user would
> need to mark the "origin cube" to define reference points. After this
> stage of point-and-mark is done, the program calculates the 3D points
> for each of the photographs and uses statistical analysis to
> determine the most likely position of a point on the model.
>
> I'm not looking for depth-of-field algorithms to produce dense 3D
> models; just the points, so I can model, say, NURBS surfaces using
> those reference points. I've looked at some sites offering some image
> rectification and affine transformation with projection, but they're
> all built for dense 3D reconstruction.
>
> Here's one example:
>
http://www-sop.inria.fr/robotvis/personnel/qdelam/ArticleFG98html/Article.forhtml.html
>
> It concerns moving a modelled hand to fit the depth-of-field produced
> by the program.
>
> Any information regarding these kinds of calculation? I guess I'm not
> alone on this subject, who hasn't thought of simplifying modelling
> to, actually, child's play? ;)

http://mi.eng.cam.ac.uk/research/vision/photobuilder/aim.html

Might be useful.

Post a reply to this message

> http://mi.eng.cam.ac.uk/research/vision/photobuilder/aim.html
>
> Might be useful.

Yeah, I took a look at that... but haven't gotten far with that, since
there's no concrete mathematical model in place, only a few sketches.

Thanks anyway, I'll keep looking.

Post a reply to this message

Klemen wrote:

> 1. the real-world model containing marked control points for surface
> extrapolation

I've looked into this stuff a bit... isn't that the dream? Take a few 
photographs and then spit out a model? :)

Anyways, what I've seen done is a grid of laser light points is 
projected over the object. Using a grid of laser lights means that the 
points are always projected with known distances between them to aid in 
the calculations...

I can't find the page, but some guy was also trying to build his own 
homebrew digitizer. I think it was mentioned on slashdot if you'd like 
to search for it (I can't at the moment). Basically the object was 
rotated on a pedestal while a laser line was scanned vertically over it 
and used as a contour shape to be pulled from the captured images.

The thing I'd REALLY like to see (and wouldn't it help with this too?) 
is to be able to take a pair of stereo photographs and then have an 
algorithms spit out a depth of field image... They're doing it for the 
mars rovers, but...

Arlo

Post a reply to this message

Klemen wrote:
> Okay, not really talking about dedicated 3D input devices. What I mean to
> ask is, has anyone thought of or perhaps even done modelling with
> multiple-view camera setups?
> 

Check out Paul Debevec's page:

http://www.debevec.org/Research/

It sounds kind of like what you want, but I believe it may not work for 
all models.

Derek

Post a reply to this message