I don’t know the accuracy of the P4, but I would guess it’s better than the default 10. I’ll try running it with 5. We’ll find out if my machine can handle those parameters.
Thanks for the suggestions!
1 Like
If it is a single flight, or a flight within a short window, you can safely turn that value all the way down to 0.7, maybe as low as 0.2, because it is (in effect) for small windows of time a GPS consistency measure, not an accuracy or repeatability measure. I think the defaults should be smaller when there is no metadata, but that might take some tuning to read the range of time stamps and adjust the default gps-accuracy values based on the range of times.
2 Likes
Good to know. I thought the 10 was an elliptical geospatial error.
1 Like
Well, it is. But how much change in error do you get in a short time? It turns out: not much (usually).
2 Likes
Argh. Failed after 14 hours. I’m going to try it again with the rolling-shutter flag.
1 Like
Maybe PC-rectify: true would help
1 Like
Failed again with rolling-shutter true. I set the matching neighbors to 3 just to see what would happen on a separate test and that failed as well.
1 Like
Any idea what you’re failing on?
Same error as my original post - selecting 0 images for the densifying.
1 Like
Did you ever upload this to my server? I didn’t see it there, but there are a lot of datasets… .
2 Likes
I didn’t, but I can do it later tonight. I’ll name it “SuBB_oblique”.
2 Likes
Yeah, it does look pretty good. The strange thing is I ran the project on my home computer using the same settings and it failed after 2.5 hours with the same error. My machine is considerably slower because I’m only using 4dimms out of 16. And, another difference is I use the --gpu upon startup. So, I’m rerunning the project now without the gpu and added another 4dimms yesterday. I also enabled IOMMU in my bios, which I have yet to fully understand how that might affect WebODM/docker-compose but I thought it was worth a shot.
1 Like
Ahh, GPU vs. CPU. Those are totally different approaches under the hood. This explains much.
1 Like
OK, It ran successfully without the --gpu. So why is the gpu causing it to fail? Other stats: It took 3:06, but the GSD was 1.03 vs the 2.77 on the spaceheater (why would the GSD be so different with the same parameters?), and mine produced 55mil pts vs the 60mil from yours.
1 Like
Are you fully up-to-date?
1 Like
It’s a great question. While we have seen differences between GPU and CPU, I’m not aware of any known outright failures. This might be a good test dataset if a bug report emerges from this investigation.
2 Likes
I’ve got 1.9.16
1 Like
Update: the nadir + oblique failed again on my machine using only CPU. I’m running it now on cranky with the same settings as the oblique.
1 Like
Good. It’s really cold in Cleveland. We’ll use the extra heat in this old house.
2 Likes