Workflow validation using scale bars and drone-footage

[wizprod]

Hi all,

I have found a workflow that seem to produce millimeter accurate measurements, but I am but a rookie at Photoscan, so I will need some help in verifying my workflow so far. Here's a rough idea:

• Aligne all photos from the drone, using generic and reference preselection
• Place control & check markers (low accuracy, but high quality pinned) and scale bars (very high accuracy)
• Disable reference for all cameras
• Optimize Cameras
• Measure!

No GCP GNSS are available, so the control points have coordinates derived from unprecise references from the drone, hence the low accuracy setting for these, to make sure the scale bars are weighted higher.

[SAV]

Hi wizprod,

You should use highly accurate reference information in order to get a reliable estimate of your model's accuracy, for example by using ground control points that have been accurately surveyed using an RTK/PPK GPS or a total station.

Moreover, you have to distinguish between relative and absolute accuracy. For example, inside your model length measurements might be very accurate because you used a scalebar from point A to point B to 'optimize' your point cloud. The length of the scale bar and other length measurements might be OK, but that does not mean that point A and point B have accurate location coordinates.

One needs to be very careful when relying on GPS reference information from 'standard' UAVs. In the case of UAVs from DJI, there is a large Z (altitude) drift over time, which leads to tilted models if no other reference information is used. Here is some more info (one of my previous posts): http://www.agisoft.com/forum/index.php?topic=8306.msg41941#msg41941

When it comes to accuracy in a SfM workflow, there is a large number of things that potential add to the 'error budget':
1) Image quality
2) Scene/object texture
3) GSD
4) Image network geometry
5) Camera calibration
6) Reference quality & quantity
7) Processing settings

I think that your extremely good 'accuracy values' might not reflect reality because all reference information has been used as a constraint to optimize the point cloud (during the bundle adjustment), which means that the algorithm tries to minimize the error between all of them. If there are not check markers or check points, it is not really possible to tell how accurate your model is. 

I am also a little confused by your post. First you say that you placed control markers and then you mention that no GCP GNSS (not sure what that is, I guess you mean GCPs that have been surveyed with a GPS) are available. So how did you measure these control markers then?

All the best.

Regards,
SAV

[wizprod]

Hi SAV,

Thank you for your response. As always much appreciated.
I should start out by correcting that this flow is intended for volume measurements, where I don't need the absolute / high precision, but high accuracy.

The faked GCPs are derived from the imprecise drone GNSS, and then given the same low accuracy as the drone GNSS readings. I am under the impression that if I don't specify coordinates here, the markers only serve as checkpoints, and does not affect the the accuracy of the model.
So I set 50 m accuracy for the GCPs and the Camera Refs, and then 20 mm for Scale Bars, where I have quite precise measurements over up to 30 meters distance. That will of course give a large error on the markers, but given the scale bars are precise, it seems that measurements on the resulting model are more consistent ?

[Dave Martin]

Wizprod,

You cannot use 'derived' GCPs whose co-ordinates are extracted from a reconstruction to then assure quality of such a reconstruction.

That’s a  bit like saying:
"My foot fits this shoe that say’s it is a size 10, so my foot is size 10. When I put my foot back in the same shoe it fits perfectly, so it proves that the shoe is actually a size 10."

Dave