Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Discrepencies between colmap in Docker (v3.10) and colmap from ubuntu repository (v3.7) #2518

Open
romaintha opened this issue Apr 16, 2024 · 5 comments

Comments

@romaintha
Copy link

Hello,
I have been running a reconstruction locally on my ubuntu which uses Colmap installed from Ubuntu repository (colmap version 3.7). I am able to get a good reconstruction of my scene.
Using the same parameters, on the colmap docker image (tag latest, colmap version 3.10-dev), the mapping stage does not output the right reconstruction at all.
At first I thought it might be due to some of the stages run on GPU, so I tried to disable GPU everywhere, but it did not help.
Is there anything that is dramatically different between the 3.7 version and the 3.10?

@sarlinpe
Copy link
Contributor

Similar issues have been reported in #2427 and #2271. @ahojnnes we'll need to look into this soon.

@romaintha if you have time, it'd be good to isolate which stage causes the issue, for example by running of 3.10 with a database (features and matches) generated by 3.7, and conversely.

@romaintha
Copy link
Author

@sarlinpe , I did a bit of testing and here are my findings.

Full (feature extraction, matching and mapping) 3.7 (ubuntu 22.04) reconstruction
image

Full 3.10 (docker latest) reconstruction:
No sparse reconstruction at all

Feature extraction 3.7 and matching + mapping 3.10 (docker latest):
image
It looks decent but if I look in details, the results is way poorer than with the full 3.7 reconstruction since some camera extrinsic are quite wrong
image

Feature extraction 3.10, matching 3.7, mapping 3.10:
image
This one looks very similar to the full 3.7 reconstruction.

So according to these tests, it seems that doing the matching in 3.7 solves the problem .

One more question: is there a way to get a built docker image with 3.7? I can't find the docker tag with 3.7 sadly

@SharkWipf
Copy link

SharkWipf commented Apr 19, 2024

In the Nerfstudio discord, I've seen several cases reported that point at COLMAP 3.8 (so not 3.7) as being the last stable version, with significant regressions on 3.9.
In several cases, successfully found poses go down from like 200 on 3.8 to 2 on 3.9.

One more question: is there a way to get a built docker image with 3.7? I can't find the docker tag with 3.7 sadly

@romaintha
While not a direct answer, and not directly applicable to docker unless you make your own container image, you can still install 3.7 and 3.8 from conda-forge.

@ahojnnes
Copy link
Contributor

@romaintha Is it possible for you to share the input images so we can investigate the issue?

@ahojnnes ahojnnes self-assigned this Apr 20, 2024
@romaintha
Copy link
Author

@ahojnnes I can't share the data publicly. If you give me an email I can send you a way to download the data

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

4 participants