NodeODM parameters/ workflow options for reducing point cloud noise

thadwald · 4 May 2020 22:59

I need some tips on how to get a less noisy point cloud. I think I am getting way too many points that are just noise. I would like to try and reduce them, while at the same time holding actual data.

Some specific questions:

Which ODM parameters deal directly with the data/noise ratio in a point cloud?
Does image pre-processing make a difference? My current workflow involves running the photos through Lightroom in order to fix some exif issues with the DJI cameras. Does anyone have any experience with the effects of sharpening and contrast?
How important is lens calibration? I have been using “auto” in the ODM settings. Can I improve the noise by calibrating the lens? How can I tell whether my workflow is calibrating the lens?
What about creating a corrected rectilinear lens profile, the likes of which is used in other photography fields? Would this kind of correction yield better results or is it redundant with ODM camera correction methods?
I am shooting with a 110degree lens (Laowa 7.5 mm, Zenmuse X5 camera) does a wide angle lens make calibration more important?

Thanks

smathermather · 4 May 2020 23:45

Not knowing much about what you are seeing specifically, the first thing I would try is using --camera-lens set to fisheye first. 110 degrees is pretty wide, and auto might not be adjusting correctly.

Auto is likely to work better in the near future for wide lenses, see Yann’s ne here:

github.com/OpenDroneMap/ODM

Fixed DEMs orthophoto misalignment bug

OpenDroneMap:master ← pierotofy:gcpoffset

opened 06:51PM - 15 Feb 19 UTC

pierotofy

+22 -14

This PR should fix https://github.com/OpenDroneMap/WebODM/issues/560 and https:/…/github.com/OpenDroneMap/WebODM/issues/549. It's a problem that has come up on the forum also at https://community.opendronemap.org/t/webodm-vs-odm-differences-in-the-precission-and-quality-of-orthophotos-and-dems/1735/5 It seems to have been caused by two separate problems. - One was the decision to not include the GCP translation offset as part of the georeferenced model transform (instead to aggregate the translation value to the UTM offset). While the choice of doing so seemed to be equivalent, it turns out that when it comes to render the model to a PNG (with `odm_orthophoto`), it does make a difference for certain models (models that are at a 45 degree angle to north seem to be most affected, while "square" models are virtually unaffected, and that's why this was difficult to track down). I didn't pin-point the exact reason, but I believe it has to do with the calculation of bounds and/or the fact that the model is flipped before rendering. - Two was a bit more subtle, when we called `odm_georef` we used the 3D mesh as the model to calculate the transform matrix, but then used the 2.5D mesh to render the orthophoto. The two models are almost identical, but not exact, so after fixing the first problem a very small offset between the orthophoto and DEMs still occurred (since they were being georeferenced using two different, although almost identical transforms). I'm re-running tests on various datasets to check for problems, but so far so good. Now (good GCP, DSM and ortho alignment): ![image](https://user-images.githubusercontent.com/1951843/52877553-553c9f80-3128-11e9-99cd-ffe946941797.png) Before (odm_data_copr was particularly bad because it's at 45 degrees to north): ![image](https://user-images.githubusercontent.com/1951843/52877663-a6e52a00-3128-11e9-8fe7-0d6c8ed05768.png) After: ![image](https://user-images.githubusercontent.com/1951843/52877647-9765e100-3128-11e9-9690-e229cb664626.png)

See if that helps, we and we tweak from there.

thadwald · 5 May 2020 02:40

Here’s what I’m seeing

It just appears to me that the point cloud density is too high, so I’d like to delineate and cull the worst, and keep the best.

The Lawoa 7.5mm is a rectilinear lens, not a fisheye. It does have a small amount of barrel distortion, which I could fix but I would need to create a profile for it, as I cannot seem to find one for it online. This would entail a bit of work so I would first like to know if it would help. so…

Does it help to have photos that are near-perfect in terms of distortion?

For what it’s worth, I had about 93% overlap with these images at about 200’ altitude.

thadwald · 5 May 2020 22:34

Is there any research aimed at determining how lens angle of view and corresponding flight height (cm/pixel remaining equal) plays into flight efficiency and accuracy?