Community Forum

Author Topic: Merging different chunks flawlessly?  (Read 3286 times)

ThomasVD

  • Jr. Member
  • **
  • Posts: 99
    • View Profile
Merging different chunks flawlessly?
« on: January 22, 2015, 11:08:59 AM »
[just for the Agisoft PhotoScan Professional Edition]

Dearest wonderful Agisoft community,

I have a problem with a dataset in which I processed 462 pictures in 4 chunks. The original pictures were 'bad' data; old scans of even older dias - no .EXIF, no GPS data. Nonetheless, the 462 pictures eventually aligned quite well using marker-based 'guided' picture alignment. The 4 chunks were then aligned using marker-based chunk alignment, however there is a geometric error of about 1m between matching chunks.

So what I have now are: approximate camera positions (in local coordinate system), approximate camera calibration, 400+ manually placed markers used to align pictures and align chunks.

I'm trying to figure out whether I could use this data in order to reprocess all 462 pictures in one chunk (from align photos stage), but using the camera positions in order to somehow 'restrict' where the software looks for matches (because otherwise it will find incorrect matches in the poor quality pictures). The hope is to then create a coherent single-chunk model without above-mentioned geometric error.

What I have already tried: I tried merging the 4 chunks in one layer, then did 'export cameras' in PhotoScan XML format, made a new file with the same 462 pictures in one chunk, did 'import cameras'. However, after reconstructing the sparse point cloud, it seems like PhotoScan uses these imported cameras as 'absolute truth', and does not correct them anymore based on the matches it finds. As such the model contains the same geometric distortion as the 4 chunks merged.

What I already tried #2: I did 'export cameras' in Omega Phi Kappa txt format, created the file with the 462 pictures in one, and imported the camera coordinates from the Omega Phi Kappa file in the reference menu. I could then 'align pictures' using reference coordinate data as a guide. As such PhotoScan does not consider the camera coordinates as 'absolute', but just as an approximation. However, I could not correctly extract Yaw, Pitch and Roll from the Omega Phi Kappa file, and as such the pictures are getting projected on the wrong side of the coordinates.

If anyone has any input on how I might be able to set an 'error margin' on the imported camera data (first test), or how I can correctly export estimated camera orientation along with estimated coordinates (second test) or any other procedure for obtaining the desired results, please let me know.
« Last Edit: January 22, 2015, 11:10:46 AM by ThomasVD »

bigben

  • Sr. Member
  • ****
  • Posts: 402
    • View Profile
Re: Merging different chunks flawlessly?
« Reply #1 on: January 22, 2015, 02:07:34 PM »
A few thoughts... not sure how helpful they'll be.

1. use some cameras in multiple chunks and align by cameras rather than control points
2. Align images in one and get lens parameters from that and apply these to the others as fixed values.

The latter may help reduce the discrepancy between chunks.

ThomasVD

  • Jr. Member
  • **
  • Posts: 99
    • View Profile
Re: Merging different chunks flawlessly?
« Reply #2 on: January 22, 2015, 04:27:50 PM »
Thanks BigBen, I'll give those two a try tonight :)