Pc geometric example

[Bryce_Brooks]

Does any one have some pictures/examples that show when enabling this setting improved your model?

[smathermather]

Nothing that know where it is… . I definitely tested it with street level data and it made a substantial difference cleaning up the point cloud along the edges of objects. I haven’t evaluated it for use with drone imagery.

[Saijin_Naib]

Not live yet, but I have taken some for the docs on the flag here:
https://docs.opendronemap.org/arguments/pc-geometric/

I noticed a marked improvement in palm tree reconstruction, even.

Update boot.py

[ichsan2895]

I tested a sample. Interestingly, pc-geometric=False seems more better than True. But, I think it will be cherrypicking since I only test a dataset.

PC geometric is True

PC geometric is False

PC geometric is True

PC geometric is False

Source of Dataset = Agisoft Metashape’s images sample dataset

[Saijin_Naib]

What --feature-quality and --pc-quality did you run with?

[ichsan2895]

ups sorry, I forgot to type it,

feature-quality = ultra
mesh-ocree-depth = 12
mesh-size = 30000
pc-quality = high
use-3dmesh = true

The difference is only, the first one was added pc-geometric = True

I think it seems cherrypicking because I just only test a dataset, maybe someday I can test it more…

[smathermather]

This is representative in a sense: pc-geometric is going to decrease error, but not increase coverage. You likely had lower quality points in the facade of the building that didn’t meet the thresholds for inclusion with pc-geometric.

So: you probably don’t want to use it in this case, but I may want to use it to trim out sky pixels from ground shots, or improve point cloud quality for elevation models I’m deriving from drone imagery, but you prefer to keep as much of the point could as possible to improve your final mesh for this building facade (which looks fantastic, btw).

[ichsan2895]

Whoaa, I like your analysis. It really makes sense. So, In short, pc-geometric helps to keep points that look suitable for that geometric. So, it decrease error because removing unwanted point cloud that outside that geometric-threshold.

[smathermather]

Yes, exactly!

[Bryce_Brooks]

So in a sense it’s a filter

[smathermather]

Yes. Although now that I’m thinking about it, the filter might apply to the mesh not the cloud. If I remember correctly, the process removes faces from the mesh that don’t fit a reverse ray trace of the data.