Forum

Author Topic: Face scan, very different result from different person, pics seems to be good  (Read 5014 times)

Tyler J

  • Newbie
  • *
  • Posts: 34
    • View Profile
hi,

recently I made a 3d face scanner based on 21 point and shoot cameras.
21 coolpix l27, this was a bad choice, I did not know that low end cameras
did not allow settings such as ISO or shutter speed to be set.
Canon camera would have allowed it through firmware hacking and unfortunatly
there is no community based around nikkon point and shoot camera firmware hacking.

I modified the cameras so they could be triggered and made my own power adapters.

That being said, I'm having good result using the "auto" mode force to macro
which in the end always uses the macro scene.
In this mode the white balance can be set, i've set it to "flash" as I'm using cool white lighting.
however shutter speed and ISO is choosen by the camera. I lowered the exposure to -0.7
and each camera takes a shot with a 1/25 shutter speed and and ISO value betwen 80 and 120.

lowering exposure to -2 ensure an ISO of 80 but the lighting required to get a good picture
will hurt the eyes of the subject (lighting is constant)

pictures are being taken in the dark, only the lighting of the scanner is present,
it's composed of 7 30cm cool white led strip with transparent paper to diffuse the light.
CRI seems to be good enought, however I can see reflection on the subjects (is this a problem?)

Here is the scanner, heigh can be set and distance from the head too.




Here is my problem, subject1 is "perfectly" modelized, I can only align half of the pictures
for subject 2 so I get half a face.


subject 1

subject 2

Why this difference? here is a link for the hi-res pictures of subject 2 :
dropbox link
As you can see 19 of them are sharp, 2 of them are not focused on the face (one is focused on it's shirt, one on it's hairs)
removing the 2 bad pictures does the same thing, only half of the face is aligned.

what do you think about the orientation?
Vertical may be a bad solution, at first I tought that by maximizing the
face area in the pictures would generate more data.

But in the end only 1 third of the face is in focus.

by putting them horizontally I get only a piece of the face but it's all in focus.

Any idea why?

Thanks

Ps : if someone here has a service manual please PM me :D I'd love to make a board capable of intercepting communication from the microcontroller in order to set ISO and shutter speed.
« Last Edit: May 14, 2014, 01:12:56 PM by Eildosa »

xiu2014

  • Newbie
  • *
  • Posts: 24
    • View Profile
I think the card camera photos need correction, the photo is deformation? If convenient, you can put the source file sharing.

Tyler J

  • Newbie
  • *
  • Posts: 34
    • View Profile
you think lens distortion is the problem?
the source files are here : https://www.dropbox.com/l/4kDbG0VVt9CWAsgASxuqpq?

James

  • Hero Member
  • *****
  • Posts: 736
    • View Profile
Part of the problem is lack of exif data in the images, another problem is you are on the very limit of required amount of overlap in some places, this could be improved just by moving the cameras slightly further from the subject.

I forced an alignment by adding a few strategic markers in places, and adding some camera parameters (not sure how accurate they are) for pixel size and focal length, and got this within about half an hour.

Photoscan can do a better job of aligning sometimes if you add some markers to a couple of aligned images, and then again the same markers in a non aligned image, then right click the non aligned image(s) and select align selected images. It is very powerful! I guess this only applies to professional version however.

I put the psz file on dropbox for you to inspect anyway - https://dl.dropboxusercontent.com/u/28695776/test.psz

I must state i only have one camera myself, and generally only scan large structures, never faces or people, so am no expert in this particular field!

Tyler J

  • Newbie
  • *
  • Posts: 34
    • View Profile
Amazing.
However I do not own the pro version so I guess it's out of my reach.

After several other experiment it seems that subject 2 has a larger head than 1, so moving backward the camera (2cm) solve the problem.