Poor processing results after update

Latest WEBODM on Ubuntu 20.10 (docker)
Chromium
Dataset from Sequoia 4 monobands 1852 pics (463 views)
Processing with preset “Multispectral” leads to very poor stitching and georef result (image)

Bildschirmfoto von 2020-11-07 20-13-58

Dataset from the same area from Mapir OCN works fine.

Logfile
Results

Hubertus

2 Likes

Did it look better before updating to the latest version?

Also do you get a better result if you turn on —use-hybrid-bundle-adjustment?

1 Like

Yes, before updating the results looked perfect.
Bildschirmfoto von 2020-11-08 07-36-21

The test with use-hybrid-bundle-adjustment is running

I also updated ubuntu to 20.10 - could that be a reason for the problem?

3 Likes

Just cancelled the run with use-bundle-block-adjustment after 5,5 hrs - usually the dataset took 1,5 h to finish.
logfile
Next try with a smaller subset of the dataset that I can share.
BR Hubertus

The use-hybrid-bundle-adjustment will certainly increase runtime. I would let it run to completion to see if it improves results.

1 Like

Another thing to try: does it work if you change camera-lens to perspective?

Changing camera-lens to perspective or fisheye makes no difference. Now trying use-hybrid-bundle-adjustment with the smaller dataset.
Perhaps someone else can try to process the dataset?

Hubertus

use-hybrid-bundle-adjustment also makes no difference. Sharing the results here.
Hubertus

Using the images of just one band (NIR) leads to a perfect result.
Bildschirmfoto von 2020-11-08 19-29-41

3 Likes

I found this bug in OpenSfM https://github.com/mapillary/OpenSfM/pull/671 which may or may not be the cause of this issue also.

Once the docker images have rebuilt (wait a few hours), I’d recommend updating and trying once again with default settings. See if it improves then?

1 Like

Unfortunately the result is still the same. I found a few topics in this forum about a “bowl effect” that looks similar to my problem. Here is a screenshot of the 3D-Model.
Bildschirmfoto von 2020-11-09 09-58-17

2 Likes

Maybe try pulling the update again. I saw this too before update:
image

In general, the posts regarding lens distortion prior to the Brown-Conrady camera model becoming default are an unrelated but similar problem, at least in symptom, if not in underlying structure.

1 Like

Made a fresh pull (cleaned all docker images before), but with no success.

I’m seeing the issue still too:
image

I’ll run one more update to be sure, but I think we have a regression in OpenSfM in addition to the one Piero caught.

edit:
Now that I say that, I have 6 nodes, and I’m not positive I updated the one this ran on. Weeeh! Infrastructure.

2 Likes

Rerunning now. @Hubertus_Paetow-- when you reran after update, did you switch it back to either brown or auto for the lens type? I think Piero had you test perspective, and if it was still perspective in the rerun, that would introduce another very similar issue.

1 Like

Changing lens type made no difference, so I switched back to auto. The dataset definitely worked well 3 weeks ago and processing one band a time also works. But for my usecase it is necessary to have all 4 Bands in one orthophoto.

1 Like

For anyone who wants to test: Here is the small dataset.

2 Likes

What parameters are you using to run? I’ll put this in my queue.

Testing is done with default parameters.
When calculating whole fiels maps i use a combination of Multispec and hires settings: min-num-features:12000, radiometric-calibration:camera, depthmap-resolution:1000, ignore-gsd:true, texturing-data-term:area, texturing-skip-global-seam-leveling:true, optimize-disk-space:true

2 Likes