Would this work? Video, turntable solution

[Xilstudio]

Would this work?
I have a wooden lazy susan, rigged up with a simple pulley.  You pull the rope and it unwinds as you pull.  works, cheap, easy.
What if placed a human model on this, set up for say 4 or 6 video cameras set for HD.  you record the rotation... you export the films as jpg sequences, choose the frames you want.
I know there would be quality loss since I would not be using 70 DLSRs, but would it function?  I am trying to a balance between speed, cost and quality.

[Infinite]

Would this work?
I have a wooden lazy susan, rigged up with a simple pulley.  You pull the rope and it unwinds as you pull.  works, cheap, easy.
What if placed a human model on this, set up for say 4 or 6 video cameras set for HD.  you record the rotation... you export the films as jpg sequences, choose the frames you want.
I know there would be quality loss since I would not be using 70 DLSRs, but would it function?  I am trying to a balance between speed, cost and quality.

You really need to record uncompressed video. The video compression on most consumer cameras is very poor. Even at HD, also 24fps is undesirable. The only benefit are the lens ranges available for DSLR's

[RalfH]

As has been said before in this forum, video is a poor source of data for structure-from-motion approaches unless you are working with a high-end camera which (1) has very high true shutter speeds and (2) allows you to save uncompressed video.

Also, in the described setup it would be extremely difficult to find synchronous frames in the 4 to 6 video sequences. If there is any movement of your model (except for the rotation), everything will be messed up. Living people breathe, for example.

[andy_s]

Does anyone know of / comment on available [to mere mortals] "non-rigid registration" type software that may help ? eg create 3 high quality photoscan pointclouds [distributed in time (and therefore some deformation)] then "stitch" them together with said software ?

See - http://www.youtube.com/user/haoli81 - only if you can spare 64minutes 

[RalfH]

You mean this stuff? (http://www.mpi-inf.mpg.de/resources/deformableShapeMatching/EG2011_Tutorial/slides/2.3%20Non-Rigid%20Registration.pdf) Ask the author of that presentation.

But why would you want to do this? If all you want to do is create a 3D model of a living person, using multiple synchronous cameras is much easier, at least if that person comes into your studio and agrees to have such pictures taken.

[andy_s]

Thx RalfH - that's a good slide summary of the video. It's an economic comparison i'm considering: less cameras v more time/deformation xtra software... 

[RalfH]

andy_s,

"time is money" could be a counter-argument. My opinion is that you'd be much better off using a number of low-cost (even used) point-and-shoot cameras with which you can shoot synchronously (see, for example: http://stereo.jpn.org/eng/sdm/index.htm). And  20 used point-and-shoots are probably cheaper and have higher resolution and better image quality than 6 good video cameras.

[andy_s]

Appreciate that Ralf.

My lower limit is 18MPix

The compare I have in mind is [say] 60 cam instant capture v [say] -

16 cam [arranged in 4 groups of 4 at 22.5 degrees to each other] then

For i=1 to 4
capture point-cloud, rotate turntable 90 degrees
next i.

[say] 10 second total capture.

Stitch 4 [overlapping] point clouds together with said non-rigid registration software.

Not sure the concept is sound but i'd like to investigate further. 

[RalfH]

andy_s,

don't know what you're aiming at - in the beginning youtalked about HD video cameras (which have approx. 2 MPix), now you say your minimum is 18 MPix?

If whatever it is that sits on your turntable is alive (e.g., a person), you'll have problems. Mislaignments and/or ugly model erros. With anything that lives, synchonisation is the way to do it. If it is just objects, that setup might work. But make sure that your rotation angles are not multiples of your angular camera spacings or you'll have identical positions with different cameras (i.e., no additional information).

[andy_s]

Hi Ralf,

it was actually Xilstudio who referred to HD Video.

It is the "problems" that you refer to that I would like to investigate whether the "non-rigid registration" software would greatly [and hopefully economically] reduce.

I appreciate your further comments.

[RalfH]

andy_s,

sorry, had overlooked that...

[andy_s]

No need to apologise Ralf - very easy mistake to make.

[RalfH]

I only know the "non-rigid registration" stuff from a quick search after your post, but to me it sounds complicated and work-intensive (and potentially error-prone), so personally I'd stick to a proven method for now.

[andy_s]

This maybe useful:

http://www.pcl-users.org/Different-geometeies-and-non-rigid-3D-Point-Cloud-Registration-td3766532.html

I wonder if they did get it into Meshlab yet:

http://www.pcl-users.org/Adding-Point-Cloud-Library-to-Meshlab-in-Qt-td4022240.html

[hsmith]

Andy,

I think your suggestion about using non-rigid registration to align meshes that you get from scans using less than 10 cameras is quite brilliant, as a low-cost alternative.
 
I've been corresponding with Hao Li about using his BeNTO3d software (check out http://www.bento3d.com/BeNTO3D/BeNTO3D_-_teaser.html )
He's seems pretty willing to help, but I haven't given him very good sets of data yet.
Still working on my rig.

Have you tried some of the non-rigid registration programs that have been developed for medical research?  3D Slicer seems the most user-friendly, (http://www.slicer.org/ )
but there are also a few others, like elastix
( http://elastix.isi.uu.nl/index.php. )

Keep up the great experimentation!

Harold