Author Topic: Photogrametry with few images  (Read 196 times)

2018-12-06, 19:42:26

lolec

  • Active Users
  • **
  • Posts: 69
    • View Profile
Hi,

I have an odd problem and maybe some expert here can help me get un-stuck.

I need to do the best possible 3d reconstruction of a human body based only in 5 pictures taken at 90, 45, 0 -45 and -90 degrees.  Before you ask, no, I can't take more pictures, unfortunately, as I'm using an existing dataset :(

Of course, I know the reconstruction won't be great or maybe even good, based on the extremely limited data set.

Some requirements:

-There are 200 human sets, so doing a manual 3d model for each one is not feasible, but doing some manual work is reasonable (like manual alignment, manual point matching etc)
- The 3d model should be textured too.
- The pictures have plenty of details and texture
- The pictures where captured with a filter that removes all reflections so the texture detail is pretty solid and there are no specular occlusion
- the 90 and 45 degree images are not precisely at 90 and 45, but close
- I only need to build the front of the body
- The camera FOV is known

I've done research using many photogrammetry software and could not find one that has a special "low number of images" mode. Maybe a human-assisted mode?

Hope someone here can help.

Thanks :)

2018-12-06, 20:09:32
Reply #1

pokoy

  • Active Users
  • **
  • Posts: 1256
    • View Profile
We've commissioned a few full body scans once and they took 80 images for each scan. The resulting models weren't very satisfactory and the shop wasn't able to deliver detailed models using Photoscan, the models just lacked details. The quality was better after they switched to Capturing Reality but still, it wasn't really that good. Textures however were quite good, highres and sharp, mapping was perfect, too.
I don't know if it was because of limited skills on their end but 5 images isn't a lot, and getting a good quality will be difficult. I guess you'll have to touch up manually a lot. Especially if you have no pictures from top/bottom, it may be too little to compute a useful mesh.

There's a good photogrammetry community on facebook - sorry, forgot their name and not on fb anymore so can't look for them - I think you'll be better off asking people doing this on a daily basis.

2018-12-06, 20:31:53
Reply #2

lolec

  • Active Users
  • **
  • Posts: 69
    • View Profile
I'm actually more interested in mapping that model quality or detail.

I have some experience with photogrammetry, we use Capture Reality. I know all the reasons why this won't work, but still, I've gotta keep trying :)

Thanks for the tip on the Facebook group, I will post there and see if I get lucky.


2018-12-06, 20:55:56
Reply #3

pokoy

  • Active Users
  • **
  • Posts: 1256
    • View Profile
I googled the fb group and found it:
https://www.facebook.com/groups/1439036619645915/

Good luck with the project, sounds like a lot of work.

2018-12-06, 21:24:48
Reply #4

romullus

  • Global Moderator
  • Active Users
  • ****
  • Posts: 5105
  • Let's move this topic, shall we?
    • View Profile
    • My Models
Maybe you could take generic human body model and project textures from photo. I think that would look much much better than mesh reconstruction from just 5 photos.

2018-12-06, 23:06:24
Reply #5

lolec

  • Active Users
  • **
  • Posts: 69
    • View Profile
Not a bad idea, the problem is there are 200 different subjects. Any idea on how to approach that?

2018-12-07, 00:16:17
Reply #6

romullus

  • Global Moderator
  • Active Users
  • ****
  • Posts: 5105
  • Let's move this topic, shall we?
    • View Profile
    • My Models
If camera positions are the same or very similar across the set, then maybe it's possible to set up 5 cameras rig in 3ds max and with camera mapping project photos onto the model and then bake it with RTT. I made small proof of concept scene, you can open it and see if that would work for you. You can replace fallof masks with gradient ramps or whatever suits your situation better. And change Corona to scanline :]