Full Body Scanner / aligning of photos

[Christian]

I am currently setting up a full body scanner using inexpensive 5MP camera modules. I am not looking for a perfect 1:1 scale usage of the models - models will be later printed out between 2 cm (!) and 20cm - I do not need all details - even though I love the work of INFINITE/LEE. Thank you for all inspirations!

This setup and quality works fine and the results are promising. Today we are setting up 40 cameras for 180 degree - I expect that we end up with 100 cameras (total material cost of that full body scanner will be less than 10.000€). Plan is to roll out some hundrets of that 3D photo booths around the world 

My question / problem is:

If I use the first setup up pictures (27 pcs) at once (load all into PS) and then apply a background mask, PS is not able to align the photos (Attchment 1). The cameras are completly wrong orientated. If use chunks (also with applied background mask) and put inside the first 8 pictures (head) - all pictures are alligned perfectly. If I put another 4 to that "working" chunk, again fine - photos are alingned (attachment 2). This works perfectly until some photos are not alignable (in our case the feet where the cameras are not set up perfectly at this time...) (attachment 3).

How can we solve that problem that we first have to detect the "distroying" pictures and than can build a model? I would love to put all photos inside and get the non alignable photos sorted out - but why are the photos that can obviously be aligned perfectly also disorientiated??

Any comment welcome...
Christian

[fabberlounge]

Hi Christian,

As you already know, PS is quite sensible with materials, sharpness of pictures, DOF... There are a lot of influence factors, we are running a multicamera (50xCanon DSLR) 3D studio, also for the purpose of printed mini-mes (fabberlounge.com). My recommandation would be to find an optimal model (denim, lot of well textured fabrics.. we are using a dummy with different sorts of clothes and some markers as a calibration tool. Once you have a good alignment, you can easily export the cameras and import them again for more problematic chunks(unless you dont rename your cameras, or change the camera settings or position) with that workflow, your cameras will always be perfectly aligned... I also thought about an alternative solution to the more expensive 600D's, would be good to compare he two systems... What my experience is, you still need some work on postprocessing, even with a really good scan in order to get a 3D printable file in a good quality.

all the best, Andreas

[andy_s]

... I also thought about an alternative solution to the more expensive 600D's, would be good to compare he two systems...

Hi Andreas,

I [finally )] just started some chdk / chdkptp development  see http://chdk.setepontos.com/index.php?topic=6231.msg107865#msg107865 - the shutter release period does seem to be stochastic though so dark[ish] room and long exposure seem likely.

Interested if you might share some further thought on the alternative you mention.

[fabberlounge]

Hi Andy,

good to see that you still in the game  we are also doing some landscape and building scanning, for that purpose we are using a hexacopter with a Sony Nex5 ( just for the weight, its a really good camera and not as heavy as comparable products) , we have done a lot of trials to get a robust servo for triggering ( and even focussing, using autofocus), at the end its a quit cheap and reliable solution, I was just thinking of using that remote technique for our multicamerasetup, we would be independent from CHDK... and could use any cheap camera for simultaneously triggering ( once the camera settings will not be changed.. ) and maybe a wifi system for saving photos on one computer.... its an early idea, have no time to investigate further will keep you updated,  all the best Andreas

[Alexey Pasumansky]

Hello Christian,

Can you send to us these sample problematic datasets for internal checks?
Please also post a couple of screenshots with Initial/Adjusted values from the Camera Calibration window.

[Christian]

@Andreas: Thank you for your input - yes - this will help after the complete setup to have reliable and repeatable scans. We will do that. Thank you. The main problem at this stage is that if you put 25 photos inside the workflow everything is fine - if you add just one "not alignable" photo, even the other 25 cameras are not correct anymore..

@Alexey - Of course, I will send you the dataset - I will contact you with a private message to get details where to send to.

@All: Just to make sure to point this out: I am aware that we are using VERY poor quality pictures and we are working at the edge of the possibilities. Sometimes we have less than 8000 detected points to be matched. The model is based only upon a few more than 3000 points.... I know that better cameras (better pictures) will solve all problems. But this is not the point at this moment - my problem is that the program works perfectly until there are some "bad" pictures inside - beside of sorting them out/dropping them, everything breaks...

[Alexey Pasumansky]

Hello Christian,

Could you specify the camera model used? I think that missing initial calibration values (like sensor pixel size) is the main problem for this dataset. I've just input 0.0015 mm for sensor pixel size in Camera Calibration window, so that automatically calculated focal length became about 2400 pixels and started Align Photos procedure with the masks from background applied (even they are not perfect). As a result I've got fine alignment without any problems:

P.S. Version 1.0.0 was used.

[Christian]

Of course - the camera is a Omnivision OV5647 (http://www.ovt.com/products/sensor.php?id=66)

Pixel size should be 1.4 x 1.4 µm...

and no - i have not been entering it into my config - I even have not thought about it... sorry. Will try that and do that NOW....

[Christian]

COOL! With 0.0014 x 0.0014 it works as expected!

Seems like an "beginners mistake" - I hope that I haven't distrubed too much and maybe some other beginners learned a lesson ;-)

Thank you all!

[mala]

A Raspberry Pi rig ?

Cool, wondered when someone would do this 

[Christian]

Yes.... They are Raspberries...

we planned to use Canons - but we did not want to run into too many syncronizing problems. Therefore - and as our background is tcp/ip networking and other it stuff - we decided to have something "controlable" within our hands. Raspberry PI is (beneath Arduino and similar) a good solution for that... then we decided just to add some camera modules to have some fancy 3D HD video as a add on.... well that was the day we decided to go a completly other way and focus on the cameras and build a "cheap" 3D photo booth...

[mala]

When they first mentioned the possible camera module I thought it would be great to have an array of them but possibly with a bigger sensor.
Is the sync working well?  And you have a managed system to download all the images to a file system on one PC?

Could you imagine if there was a networkable PI version of Photoscan,you could then use all the PIs as a "render farm" to help process images to build the mesh!
...in my dreams!

[andy_s]

.... its an early idea, have no time to investigate further will keep you updated

Good stuff Andreas...keep at it 

[Christian]

@mala - the RaspberryPI is a network device - this means - downloading and managing is implemented by it's roots. It is really the other way round - you have all the possibilities with sync and timing - but have lack of optical system... we all do have our challenges  ;-)
Using RaspberryPI AND real Cameras is the perfect world - but who want's to live in a perfect world?

[mala]

Well the good thing is they are cheap and if the system scales well you can just keep adding more units.
Would be interesting to try some face scans with the rig as well.
Hope to see some of your meshes from the rig soon