Forum

Author Topic: 3D modeler  (Read 12174 times)

WzDrg

  • Newbie
  • *
  • Posts: 7
    • View Profile
3D modeler
« on: May 22, 2011, 08:34:31 PM »
I'm currently using blender to edit the generated 3D model. But somehow I currently can't match the original image with  the corresponding camera.
What 3D modeler can I use? Using which exchange format?

Dmitry Semyonov

  • Agisoft Technical Support
  • Full Member
  • *****
  • Posts: 175
    • View Profile
Re: 3D modeler
« Reply #1 on: May 24, 2011, 01:02:45 AM »
Hello WzDrg,

PhotoScan supports saving camera positions with a 3D model in 3DS, VRML and Collada formats. As only camera FOV can be saved in these formats, it is required to remove nonlinear distortions from the source photos to get the perfect match in 3D modeling software. This can be done using "Undistort Photos..." command from the Tools menu. Please make sure to select both "Center principal point" and "Square pixels" options in the "Undistort Photos" dialog.

In most cases that should provide matching with a subpixel accuracy. If you still observe large mismatch, please send the steps you are performing to import 3D model and photos into Blender so we could reproduce the problem. You can send them to support@agisoft.ru.

With best regards,
Dmitry Semyonov
AgiSoft LLC
With best regards,
Dmitry Semyonov
Agisoft

KEgg

  • Newbie
  • *
  • Posts: 34
    • View Profile
Re: 3D modeler
« Reply #2 on: June 02, 2011, 07:22:20 PM »
Hi WzDrg,

I've been using Blender for camera positioning since day one of finding Photoscan, and currently working on a mesh conformation and alternate texture creation tool. I've been using my own custom build of Blender for linux, using Collada exported mesh.

I export two set of images from Photoscan, one set being 'Center principal point', the second without, so the lens distortion is seen full frame. In blender, after changing the image resolution to match that of the photos shot by the cameras, I bring in one of the 'Center Principal' photos and display it in the background of its corresponding camera. Often you'll find you may have to adjust the FOV to match that found in the exif information, or adjust by site till it matches. From here I create a new 1k squared blank image within blender, and create a new UV projected from the corresponding camera. Swap out the blank 1024x1024 image with the 'Center Principal' for that camera.

The reason I export two set of images, is because you'll have to create a new style of image so that the UV can repeat properly. See attached image. It would be great if this was a feature of photoscan.


EricM

  • Newbie
  • *
  • Posts: 12
    • View Profile
Re: 3D modeler
« Reply #3 on: June 03, 2011, 04:13:53 PM »
Hey Kegg, sorry to hijack the thread, but that's a very clean mesh that you have here... Was this straight out of Photoscan or is this already a cleaned-up version ?

My I dare to ask what setup/settings you used ? (number of photos, resolution, meshing/smoothing parameters...).

My early test, while encouraging, are nowhere as clean as yours...


KEgg

  • Newbie
  • *
  • Posts: 34
    • View Profile
Re: 3D modeler
« Reply #4 on: June 04, 2011, 12:59:51 PM »
Hi Eric,

The final result is mildly cleaned up, and have removed a lot of the high frequency noise of the resulting output from Photoscan, but was literally a few second job using sculpt tools to smooth it down without loosing its accuracy.

I pretty much use the default settings that Photoscan has when you first start it up after install. The real trick is getting the sharpest image possible with a good amount of depth of field within the frame. With the right amount of light and perfect image quality, the results get smoother and smother within the package. Also the use of masks can really improve alignment. I only use exact/high for reconstruction. Also, I'm not using a single camera to capture with and have created my own multi camera rig so I'm able to capture facial expressions in an instant. This example was done using three cameras that create two stereo pairs.

I've been using Di3D systems for the last 7 years, now using alternate tech, have treated Photoscan in a similar manner. Knowing Di extremely well, I know what those guys do to make their scan data look good, and smoothing the resulting mesh (noise reduction) is part of their pipeline. Hope it becomes part of PhotoScan at some point.

Currently waiting for the next Photoscan update, as Dmitry mentioned that there should be a fix to the exact reconstruction method so the split that occurs is no more. This gives issue when it comes to clean up.

Attached an image of the result within Photoscan, which is highly comparable with results from Di. Only if the software had the ability to do noise reduction to the final result. No masking used for this, and no holes filled.



« Last Edit: June 04, 2011, 01:09:36 PM by KEgg »

EricM

  • Newbie
  • *
  • Posts: 12
    • View Profile
Re: 3D modeler
« Reply #5 on: June 05, 2011, 11:01:04 AM »
Thanks for the answer and the picture.

That's more in line with the later results I now have, except that I'm using "smooth" reconstruction usually because I can't seem to  fill the holes in photoscan after an "exact" reconstruction (yet... ).

I'm using Zbrush to smooth the last bits and reproject the detail on a clean topology. I also noticed that reflections on the skin can really make or break the results. I'm going to invest in some light make-up to matifiy the brightest parts of the forehead and nose.

Cheers,

Eric

« Last Edit: June 05, 2011, 11:03:26 AM by EricM »

KEgg

  • Newbie
  • *
  • Posts: 34
    • View Profile
Re: 3D modeler
« Reply #6 on: June 06, 2011, 02:45:20 AM »
Also use light makeup this end, and polarizing filters for the cameras. Haven't used then on the example above, and I'd removed the soft boxes from my flash guns, as it was an experiment in shadow and specular pickup. Love to see the results you're getting Eric

ta,

Mike



EricM

  • Newbie
  • *
  • Posts: 12
    • View Profile
Re: 3D modeler
« Reply #7 on: June 06, 2011, 01:21:25 PM »
Here's an example.
http://imageshack.us/photo/my-images/402/headprojecttest.jpg

I shot a 180° head scan from 10 photos, shot in daylight with no makeup or polarized filter to .

The reconstruction (smooth/High/filter 20%) was exported to Zbrush as an OBJ.

In Zbrush I cleaned it up and aligned it with a head model I already had with good UV and topology and then projected the geometry onto the clean mesh.

It was my first attempt, and it did a very good job overall, except in the places where the target model had overlapping geometry (most obvious in the lip corner region). I need to fix the target model a bit, and then it should be alright.
« Last Edit: June 06, 2011, 06:17:22 PM by EricM »