Forum

Author Topic: Huge marker error (over 1000 KM)  (Read 2645 times)

trigauts

  • Newbie
  • *
  • Posts: 4
    • View Profile
Huge marker error (over 1000 KM)
« on: July 24, 2020, 12:59:45 PM »
Hi,

I am trying to capture a WWI bunker in Agisoft Metashape but I ran into an issue regarding marker errors (m).

I used natural markers to align different chunks I made, some of them with different cameras (e.g. iPhone for window details, drone high up for the natural environment, sub-par lit interior with a DSLR). I manually measured a distance between two of the markers in real-life to scale my model (I don't have a Total Station). The chunks seem to be aligned perfectly and the scale also holds up very well.

However, when looking at the reference pane, the error in metres is completely through the roof (some errors are several thousand kilometres. I have no idea how to resolve this issue. I tried optimizing cameras, (un)checking markers etc. but nothing works. I also took great care of only picking pictures where the natural reference points are clearly visible before adding a marker. There also seems to be something wrong with the estimated camera locations, when looking at the report. The model looks great and it is correctly scaled, however, for reporting I need to have an acceptable error.

How should I get the errors to acceptable levels again? I attached my report and som screenshots for clarity.

Thanks a lot in advance.

Thomas

James

  • Hero Member
  • *****
  • Posts: 748
    • View Profile
Re: Huge marker error (over 1000 KM)
« Reply #1 on: July 24, 2020, 01:16:38 PM »
In the reference pane, you can right click on the markers with the high errors, and select 'show info' (i think) and i think this will then list the error in each image for that marker. this could help you track down the problem if only a small number of images have a very large error.

Alternatively, it looks in the report as if a large number of images were taken from identical positions (the ones with the big ellipses). If the particularly bad markers 1-4 only appear in images from one of those single vantage points then they will actually be fairly useless and the errors completely meaningless.

trigauts

  • Newbie
  • *
  • Posts: 4
    • View Profile
Re: Huge marker error (over 1000 KM)
« Reply #2 on: July 31, 2020, 06:10:31 PM »
Hi James, thank you very much for your reply.

I followed your advice and right-clicked on the markers with high meter errors to "Show info...". Here the errors were only shown in pixels. Although the pixel error was acceptable I think (all below 0.5 px), I removed markers where the error was higher than 0.1 px. This only made the overall pixel error smaller and did not change anything about the huge meter errors. "Update Transform" and "Optimize Cameras" did not help either.

Indeed, when looking at the "Camera Locations" in the report, it seems that a lot of pictures were taken from the same position, but there's the rub, they weren't. I attach an image of a comparison between the "Survey Data" and the "Camera Locations" in the report. The cameras seem to be correctly identified in the "Survey Data" section, but the report bundles all camera locations together in the "Camera Locations" section for some reason.

I circled around the building with a drone, taking one picture per camera location. I never rotated the drone around its own axis to take several pictures from the same location, as the "Camera Locations" estimation wrongly suggests.

Any idea what is going on here?

James

  • Hero Member
  • *****
  • Posts: 748
    • View Profile
Re: Huge marker error (over 1000 KM)
« Reply #3 on: August 03, 2020, 06:00:33 PM »
i think i get it now.

the 4 positions where it shows the cameras bunched together are the 'source' coordinates for those images. for whatever reason there is some rounding error which 'snaps' them to those 4 positions. perhaps it has a built in gps which is only precise to a few 10s of metres or something, which also explains why it's rounded differently in x (eastings) to y (northings).

the other 'ends' of those ellipses show the 'estimated' camera positions which matches with and are correctly shown in the 'survey data' section of the report. i've overlaid the two plots attached to show that.

you should go to the cameras section of the reference pane and 'clear' the coordinates for the drone pictures. (select the rows and right click-clear).

the list of markers you showed me... are they the source coordinates or estimated coordinates? if estimated then perhaps the above will sort those out. if there are the source/input coordinates for the markers then that is also contributing to your problem as there is such a discrepancy in the locations given for the first four markers compared to the other six.

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14813
    • View Profile
Re: Huge marker error (over 1000 KM)
« Reply #4 on: August 08, 2020, 03:39:54 PM »
Hello Thomas,

For me it seems strange that the camera reference is measured in WGS84 system (Lat/Long format), whereas coordinate system of the markers (according to the screenshots) is set to Local Coordinates (in meters), which cannot be true according to the values used.

Also  the coordinate information for the following markers is clearly incorrect: point 1, point 2, point 3, point 4, scale 1, scale 2.

I would suggest to set the system as WGS84 for the chunk in the Reference pane settings dialog (do not enable individual systems for markers and cameras), check on only markers with the correct coordinate data in the Reference pane (uncheck the markers with the labels mentioned above), uncheck all cameras in the Reference pane and then press Update button. It should be performed prior to the optimization, as the optimization of the camera alignment with the incorrect coordinate data may affect the tie point cloud considerably.
Best regards,
Alexey Pasumansky,
Agisoft LLC

trigauts

  • Newbie
  • *
  • Posts: 4
    • View Profile
Re: Huge marker error (over 1000 KM)
« Reply #5 on: August 09, 2020, 07:39:13 PM »
Dear Alexey, dear James,

Thank you for your kind reply. I have tried both your suggestions but I cannot seem to get rid of the high errors (see picture 1 attached). Completely deleting all markers, deselecting the DJI cameras, updating and optimizing, does lower the camera errors somewhat but they are still +10m and I lost all my markers in the process (see picture 2).

Just to clarify: none of the markers have measured/correct coordinates. I did not have a total station for data capture and I only manually added markers to be able to combine several chunks (tie points gave me bad results). The coordinates are estimated by Agisoft, I presume based on the GPS info baked into the DJI cameras, which I didn't turn off at the time when adding the markers.

In short, I would like to be able to do the following with low errors:

1. align/merge different chunks based on markers;
2. create a scale bar for two of my manually added markers to be able to scale the model as best as possible without measured coordinates (I manually measured the distance between several markers with a tape measure on the real-life heritage asset);
3. I do not care for the coordinate system, as long as the model is correctly aligned and scaled according to the distance I manually measured.

Do you know how I can best achieve this without having huge errors?

Should I have deselected the DJI camera coordinates from the very beginning before adding my markers? Should all cameras (also the iPhone and DSLR without GPS info) and markers be in WGS84 before aligning? Should I set the camera accuracy in the Reference Settings to a lower/higher value?

So again, what I would ideally need is a model consisting of multiple marker-based aligned chunks, where the camera locations are estimated within an acceptable error (m) and scaled according to a scale bar on the model (which I do not know the coordinates of, only the tape-measure distance between two points).

Thanks a lot for helping me out!