Forum

Author Topic: GCP's and bundle adjustment  (Read 3255 times)

jwallace

  • Newbie
  • *
  • Posts: 17
    • View Profile
GCP's and bundle adjustment
« on: August 20, 2018, 08:25:30 AM »
Hi,
This is from a project I completed a few months back.
And I was after some peoples opinions.

I was playing around with the GCP’s and updating vs bundle-adjusting…and I now have some more questions.
When I added the first 3 GCP’s, 9000, 9004 and 9010. The error was only 8mm.
When I added the rest (another 6), and ran a bundle adjustment, the error on each point increased to about 50mm.
What would you do in this situation, use only the 3 GCP’s or more?

GCP 9010 was still the outlier, at >100mm.
If I took out GCP 9010 and re-adjusted, the error reduced to in the 30mm range. But 9010, now a check point was of by 500mm. It was surveyed by someone else, but I don’t think they would have it >0.5m off.

I didn’t want to, ‘falsify’ the results by getting more accurate GCP results but poor check point results? Thoughts?
Thoughts? eg. would you include or remover GCP 9010.

Thanks,
James

Dave Martin

  • Full Member
  • ***
  • Posts: 170
    • View Profile
Re: GCP's and bundle adjustment
« Reply #1 on: August 20, 2018, 11:23:26 AM »
James,

I think more information would be really helpful, but firstly three GCPs is really the absolute minimum. I am personally not that surprised that you had minimal errors shown when using just the three GCPs -  you can always fit a plane to pass through any three points; and you can probably pull a photogrammetric surface to them provided those three points were all 'in the same field'.

The number and distribution of ground points to be independently surveyed depends on the overall size of the area, the general topography and the variations therein.  You need to record not just the points you need as GCPs, you also need a number of additional points you can use as check points - they're not used to constrain the model, but to assess the accuracy of the derived DEM.

If you can share more information about the area, topography and then a plot of the ground points and the residual errors you're seeing, that might allow more specific answers.

Dave
« Last Edit: August 20, 2018, 12:44:30 PM by Dave Martin »

jwallace

  • Newbie
  • *
  • Posts: 17
    • View Profile
Re: GCP's and bundle adjustment
« Reply #2 on: September 11, 2018, 07:46:01 AM »
Hi Dave,
Thank you very much for replying, and sorry for the delay in getting back to you.
It was a 25ha block. Relatively flat over the whole lot.
See attached image.
There were also some check points surveyed at the front of the road. The average Z error for these was 24mm.
Thanks,
James

gto234

  • Newbie
  • *
  • Posts: 42
    • View Profile
Re: GCP's and bundle adjustment
« Reply #3 on: September 11, 2018, 09:29:32 PM »
I think your results are fine within 2-3 cm for a UAV project.
Something is wrong with your photos coordinates. They are not in the same reference system as your GCP's and it is better to convert them first to the same system.
Easting= 148.xxx and Northing=-33.xxx are strange numbers for a projection. Maybe they are degrees converted to meters?

jwallace

  • Newbie
  • *
  • Posts: 17
    • View Profile
Re: GCP's and bundle adjustment
« Reply #4 on: September 12, 2018, 02:53:54 AM »
Thanks GTO.
I was happy with the accuracy overall.
My main question still relates to GCP 9010 that was the outlier, at >100mm when included as a GCP.
If I took out GCP 9010 and re-adjusted, the error overall reduced to in the 30mm range. But 9010, now a check point was of by 500mm. It was surveyed by someone else, but I don’t think they would have it >0.5m off.

I didn’t want to, ‘falsify’ the results by getting more accurate GCP results but poor check point results? Thoughts?
Thoughts? eg. would you include or remover GCP 9010.


macsurveyr

  • Jr. Member
  • **
  • Posts: 69
    • View Profile
Re: GCP's and bundle adjustment
« Reply #5 on: September 13, 2018, 05:01:15 AM »
I think you should change your marker accuracy estimate to a more realistic value than 0.005 before making a determination about GCP accuracy.

Tom