Very wrong reconstruction broken camera metadata

[grigoro]

Running a dataset with the same pipeline creates very weird results in odm.As i manage to see in logs densify point cloud not choosing all images from dataset.

14:26:57 [App ] Command line: /var/www/data/56054e5f-e562-4dd1-ace9-d296990a7f0d/opensfm/undistorted/openmvs/scene.mvs --resolution-level 0 --min-resolution 1321 --max-resolution 1321 --max-threads 16 --number-views-fuse 2 -w /var/www/data/56054e5f-e562-4dd1-ace9-d296990a7f0d/opensfm/undistorted/openmvs/depthmaps -v 0 --geometric-iters 0 --cuda-device -1
14:26:58 [App ] Preparing images for dense reconstruction completed: 221 images (688ms)
14:26:58 [App ] Selecting images for dense reconstruction completed: 159 images (170ms)

The GPS error is 36meters.

The ODM report:
odm_report (6).pdf (6.7 MB)

The log file:
odm_log (1).pdf (2.0 MB)

And the drone db image:

Any help about where the problem lyes will be great .
Best regards.

[grigoro]

Weirdly enough just reducing minimum number of features to 10000 seems to run the same dataset fine.
How just changing the minimum number of features from 20000 to 10000 can be so differentiating?I also realized that in the bad run,there were 2 camera models created for some reason.Any clue on why this happened will be very helpfull as i run the ODM pipeline fully automated.

Thanks in advance.

[Saijin_Naib]

Hard to say without the data and testing it further. Can you try with rolling shutter distortion correction if you have the sensor readout time for your data?

[grigoro]

Trying to debug,for some reason all my other datasets have only one camera created in report even running with the exact same options.Maybe this second wrong camera makes the dataset bad.I will try running the dataset using cameras.json from previous datasets.

[Saijin_Naib]

Do you mean reconstructions? If you get multiple, that usually indicates poor overlap/sidelap or other issues matching features.

[Gordon]

Clearly it is having great difficulty finding enough features when the threshold has to be reduced this low, no doubt well into the noise, which is probably why using a lower number of features gave a better result.

2022-07-01 13:44:44,592 DEBUG: Computing sift with threshold 6.698730897528758e-05

It starts here: Computing sift with threshold 0.066, then goes lower until sufficient features are found.

[YanNoun]

Hi grigoro,

Looking at the report, the camera parameter estimation went nuts for some reason. Could you share a dataset or a subset of images, so I can have a look ?

Thank you !

Yann

[grigoro]

Sorry for the late response but it were some stressfull weeks.
The images i use in the ODM pipeline get copyed metadata from initial images after some opencv operations(that clears the metadata in the result images).So there was some bug at copying metadata script in one of the images entered the pipeline,that had almost nothing as metadata and that was the problem.The pipeline recognized 2 camera models which somehow messes up opensfm algorithm.Thanks for your help.