Forum

Author Topic: View errors - orders of magnitude changed?  (Read 2251 times)

rossnixon

  • Jr. Member
  • **
  • Posts: 92
    • View Profile
View errors - orders of magnitude changed?
« on: September 06, 2019, 08:10:21 AM »
Within the last few months, a new version of MetaShape has been showing me errors in my project which are perhaps(?) 100x smaller than usual.

 I have opened some old PSZ files to check and these also have the unusually small errors. I am typically getting errors of 0.004m, whereas I would normally expect around 1-4m.

This is for aerial models. Sorry but I didn't record the date or version when this occurred.
Has anyone else seen a large reduction in their errors?

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 13190
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #1 on: September 06, 2019, 05:25:08 PM »
Hello rossnixon,

If you have the old project where you had to have several meters errors, can you share such project (without re-saving it in new version) with us?
Best regards,
Alexey Pasumansky,
Agisoft LLC

rossnixon

  • Jr. Member
  • **
  • Posts: 92
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #2 on: September 09, 2019, 02:48:43 AM »
Here is a small old (July 2017) model which would have had errors of more than 1 or 2m per point.
https://www.dropbox.com/s/1au553kwmpabq1f/Waihora.psz?dl=0

Thanks
Ross Nixon

Paulo

  • Hero Member
  • *****
  • Posts: 763
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #3 on: September 09, 2019, 06:47:08 AM »
Hi Ross,

I am wondering how your GCPs were determined...Because in the adjustment you specified a GCP accuracy of 5 mm which I doubt since you are identifying or pinning them on old scanned B&W analogue imagery.... Also the fact that the check point has an error of some 10 m proves that accuracy is not correct or points were not pinned at their correct position....

If the imagery is good (neat) and well textured you should get an average reprojection error well below 1 pixel for both images and markers.... In your case they are over 4 pixels.....


Best Regards,

Paul Pelletier

rossnixon

  • Jr. Member
  • **
  • Posts: 92
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #4 on: September 09, 2019, 07:15:04 AM »
I have never entered or changed the precision/accuracy of my GCPs. Perhaps I should.
That doesn't explain why my errors have changed to about 100x smaller. Only the Metashape version has changed.

I have changed the precision to 5m for the points I kept. That looks more like it used to be.
Is there somewhere that I can set a default precision? (It is likely to be similar for all projects).
« Last Edit: September 09, 2019, 07:25:36 AM by rossnixon »

Paulo

  • Hero Member
  • *****
  • Posts: 763
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #5 on: September 09, 2019, 08:08:34 AM »
Just go to reference settings (last icon in Reference pane) and modify Marker accuracy
Best Regards,

Paul Pelletier

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 13190
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #6 on: September 09, 2019, 02:30:14 PM »
Hello Ross,

Thank you for providing the sample project.

In Metashape Pro (and PhotoScan Pro) versions prior to 1.5.4 estimated control point coordinates displayed in the Reference pane (and the error values as well) were calculated based on minimization of the reprojection error only, without taking into account measured control point coordinates in space. So they did not correspond to adjusted coordinates from the bundle adjustment step, and were recalculated from weighted image measurements alone.
 
Starting from Metashape Pro version 1.5.4 estimated control point coordinates correspond to the adjusted values, that is, they are calculated taking into account both reprojection error and measured object space coordinates. Similar behavior is common for other software packages and also allows for better detection of the measurement errors.

If you are interested in the difference between measured and triangulated control point coordinates, you can temporary uncheck the corresponding entry in the reference pane (thus making it a check point). Please note that check point error calculation is not affected by the change mentioned above. Upon unchecking the control points you do not have to click on Update button.
Best regards,
Alexey Pasumansky,
Agisoft LLC

Ryuseiken

  • Newbie
  • *
  • Posts: 18
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #7 on: December 25, 2019, 04:45:03 AM »
Hello Alexey,

Thank you for the information :)
May I confirm the following points regarding your explanation?

Quote
In Metashape Pro (and PhotoScan Pro) versions prior to 1.5.4 estimated control point coordinates displayed in the Reference pane (and the error values as well) were calculated based on minimization of the reprojection error only

I think "minimization of the reprojection error" here means the one done in triangulation process for each point, not the one in Bundle Adjustment (optimization of intrinsic/extrinsic parameters). Right?

I believe that, in any version of Metashape so far, BA minimizes the weighted sum of "reprojection errors for valid tie points" and "estimation errors of LSE coodinates for checked markers and checked camera extrinsic parameters".

Quote
So they did not correspond to adjusted coordinates from the bundle adjustment step, and were recalculated from weighted image measurements alone.

I guess here "adjusted coordinates from the bundle adjustment step" for a control point means the sum of the observed (source) LSE coordinates and the residuals remained at the final iteration step of the optimization (BA), displayed in Chunk coordinate system. Right?

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 13190
    • View Profile
Re: View errors - orders of magnitude changed?
« Reply #8 on: December 25, 2019, 11:03:15 AM »
Hello Ryuseiken,

Unchecked marker (check point) - optimal position with respect to image measurements alone,
Checked marker (control point) - optimal position with respect to image measurements and measured GCP coordinates in world coordinate system

 
Best regards,
Alexey Pasumansky,
Agisoft LLC