Absolute accuracy with GCPs

1 Like

I wonder if one of the servers that processed the dataset hadn’t been updated (sometimes if there are lots of jobs in a queue, a server won’t get updated right away).

I re-uploaded the task and got these results:

https://wln2.sfo3.cdn.digitaloceanspaces.com/44a36cac-6c45-4754-bc98-b8d80358a407/all.zip

image

So a bit better, but still not as good as what I achieved during testing.

I should mention that during my troubleshooting, I resized the images and the GCP file by 25% to speed up processing. Doing that, plus the latest changes, gave me the results I posted.

I’m going to try to re-upload the resized images, to see if this is consistent or if I missed something obvious.

1 Like

Resized images results: https://wln2.sfo3.cdn.digitaloceanspaces.com/4b1626f5-0bea-4d88-8b58-7a7fbe421fe6/all.zip

image

So accuracy seems to improve with smaller images, suggesting that the GCP file might not be tagged precisely. I’ve found several GCP entries where the tag was not precisely in the middle of the target, e.g. GCP #1 for DJI_0443.JPG and DJI_0444.JPG:

image

I’ve manually retagged the GCPs and I’m re-uploading the full resolution task. Will share results once they are done processing.

Edited GCP file: https://testhub.dronedb.app/orgs/pierotofy/ds/w66nhtgg2rsywhwl/download/gcp_list.txt?inline=1

How did you create the GCP file @SkyView ?

2 Likes

The GCP file was created with the same program you’re showing in this post. I processed in EPSG:4326 this time. Here’s a link to my latest GCP file.

I’m a new user and, as I said, I’m running the windows WebODM lightning app. I use the default settings and I don’t resize images when I process. Please let me know if I should change any of the parameters before processing. I have some art editing programs that I’m not terribly familiar with but I’m willing to bet there’s a way to find the exact center pixel (or close enough) of the GCP targets in GIMP.

Let me know what you think. These are very exciting developments!

2 Likes

Ah cool! Yes there’s software that automates the GCP tagging process. One such example is Automatize GCP image coordinate collection - OpenDroneMap

We’d love to make these programs more accessible and user friendly long term.

So I’m looking at the images in GIMP right now. You absolutely can find the pixel locations in the grid. It looks like the zero point is the top left, like reading a book (in english :wink:). In my set of pictures, DJI_0035.JPG has four pixels that look to make up the “center” of that target. In DJI_0038.JPG it looked like two pixels were a better fit. Knowing how the grid is set up in ODM would allow me to manually enter those center points in the gcp_list.txt. If you only need 3 to 5 images tagged for each point, it’s really not much more work than using your GCP editor.
2 red center

2 Likes

Got the results in full resolution processed. As puzzling as this is, they are not as accurate:

image

:confused:

https://wln2.sfo3.cdn.digitaloceanspaces.com/d9377993-e048-4502-8551-91096600b242/all.zip

This might need further time to investigate.

1 Like

That’s how it goes sometimes. I’ve manually edited a gcp list to the pixels selected from the image editor. I’m going to run that right now. I’ll post results when I finish.

2 Likes

It definitively improve results. Thats great!
Will do more testing and see if it is consistent with different datasets.

image

1 Like

I didn’t have much luck with the last one. I’m going to try resizing the images this time. I used the default setting of 2048 and I’m using this gcp_list.txt.

We’ll see how it goes.

Screenshot 2021-09-19 055753

1 Like

hey @SkyView, are you measuring in RTK mode? What if you change to static mode?

Personally I tend to select GCP meassurement method to keep errors lower than GSD.
For instance, for a GSD greater than 3cm RTK meassurements will work; But for a GSD between 3 and 1 cm I prefer static mode. For GSDs lower than 1cm I use total station.

I must clarify that I’m not sure this will help on the model accuracy, It’s only a personal preference.

Saludos

2 Likes

I’m using two Emlid Reach RS+ units. One is mounted to a tripod in base mode. That is connected to the NTRIP (Ohio Department of Transportation). I usually average those results for 5 minutes. Once the averaging is done, I disconnect it from the internet and connect to the rover unit. The two Emlid units are connected via LoRa radio. I’m using all of the manufacturer recommended settings on both units. All of my sigmas in the survey log are within expectations. I’m happy to provide the 3 survey logs from this photo set, if you’d like them. I shot one in the NAD83(2011) CRS, one in UTM 17N, same geoid, one in Ohio South, again in NAD83(2011). Vertical datum is NAVD88 per my state’s NTRIP caster.

1 Like

With GCP Stats / Reporting Improvements by pierotofy · Pull Request #1351 · OpenDroneMap/ODM · GitHub we will now display the GCP information directly from the report/stats data of OpenSfM instead of re-computing it, so it will match the information from the report. Also working on adding an interface in WebODM to display GCP info on the map. :clinking_glasses:

5 Likes

This sounds fascinating!

I just saw that Emlid has launched an application to geotag photos. It’s still in beta but here’s the link.

https://docs.emlid.com/reach/tutorials/post-processing-workflow/working-with-emlid-studio/introduction-to-emlid-studio/

1 Like

I could set up a new survey and flight to see if this will work. It looks like Drone Deploy’s capture app does allow me to set the camera angle. I might have some time later this week if we think it’s worth doing.

1 Like

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.