Re: [rp-ml] Faux 3D Pictures

Charles Overy wrote:
> Interesting. I had not heard about the Microsoft project either
> although there have been some papers suggesting this technology for
> several years at Siggraph. The article on the Microsoft work said it
> was supposed to be announced at SIGGRAPH. If it was, I did not notice
> it. Microsoft had a booth but they appeared to be demonstrating
> Vista. Did anyone else see the MS project?

This technology has been around for about 8 years - there was a really
good SigGraph course I attended that lays out all of the details.

The Microsoft contribution is most likely just a cute GUI around the
standard algorithms.

I've written software to the SigGraph recipe - and it works well
enough providing that your photos are not taken all in one horizontal
plane - you need to take some shots from up high and others from down
low to capture all of the 3D data.

Best of all is to use video footage because point cloud tracking
works best when all of the points are close together from one image
to the next.

A lot depends on the nature of the background too - if it's very
'busy' like trees with leaves everywhere - then the tracker gets
easily mislead. If the background is too bland then figuring out
the camera positions is hard.

The algorithms out there don't work at all well with very shiney
objects because the shiney highlight moves as you move the camera
making the object very tricky to scan.

If you wanted a completely general package, you'd want the user
to be able to help out by (say) clicking on points that are the
same from one photo to the next.

It's a useful technology - but it's far from 100% robust.
Received on Tue Aug 08 04:15:57 2006