So I've been meaning to do some work on improving the orthophoto results (I've started a conversation here time ago: https://github.com/OpenDroneMap/OpenDroneMap/pull/444)
Achieving a 2.5D mesh can be done using a constrained delaunay triangulation and it's detailed here https://github.com/CloudCompare/CloudCompare/blob/579ee91594e16beba5863e2e4d15cf0729eb68a8/libs/qCC_db/ccMesh.cpp and here https://github.com/CloudCompare/CloudCompare/blob/579ee91594e16beba5863e2e4d15cf0729eb68a8/CC/src/Delaunay2dMesh.cpp
CloudCompare offers this feature built-in and has a command line mode http://www.cloudcompare.org/doc/wiki/index.php?title=Command_line_mode, however I don't think I'd want to include all of its dependencies (including QT!) just for the sake of generating a 2.5D mesh.
I could isolate just the code that takes care of the 2.5D mesh generation (and add some filtering of my own to reduce the spikes), but it will be necessary to add http://www.cgal.org/ as a dependency. PCL does not have any support for delaunay triangulations.
@dakotabenjamin what are your thoughts?
pierotofy
All 29 comments
Also cc @smathermather
pierotofy
on 25 Mar 2017
Can we also isolate just the bits of CGAL we need, or is that a fools errand?
smathermather
on 25 Mar 2017
CGAL is complicated to just use as-is, I wouldn't venture as far as isolating a single component. What are the downsides of including CGAL, besides increased compilation time and docker image size?
pierotofy
on 25 Mar 2017
Just that. We can include it and then put it on a list of dependencies to remove later. I'm happy to take that approach: we're at an add features stage for the project, so I'd say proceed!
smathermather
on 26 Mar 2017
Feel free to assign me the task.
pierotofy
on 26 Mar 2017
Can we migrate all the current PCL functions over to CGAL? I don't have any experience with it. Removing PCL could speed up complication times significantly.
dakotabenjamin
on 27 Mar 2017
Probably. The good news is that there's a prebuilt package for including CGAL in Ubuntu. apt-get install libcgal-dev.
pierotofy
on 27 Mar 2017
That is pretty good news. PCL is heavy.
I would suggest for now just getting the 2.5D work in and we can sit down and sort through the possibilities of a full migration at a later date.
dakotabenjamin
on 27 Mar 2017
Boom!
Still need to do some smoothing / filtering.
pierotofy
on 29 Mar 2017
pierotofy
on 29 Mar 2017
Awesome!
smathermather
on 29 Mar 2017
Starting to look a lot better:
pierotofy
on 3 Apr 2017
Still need to do further smoothing, but I think we're going to get free cropping along with this enhancement!
This is due to the delaunay triangulation and outlier removal which makes some nice sharp edges around the resulting orthophoto (vs poisson).
pierotofy
on 6 Apr 2017
It seems nice, how to test it ?
kikislater
on 6 Apr 2017
Checkout this commit https://github.com/pierotofy/OpenDroneMap/commit/503671487020fa4ff6dad77bdf984bb5592f5313
And run with the --use-25dmesh argument. It's still a work in progress though, expect improvements in the next few days.
pierotofy
on 6 Apr 2017
You will also need to compile the odm_25dmeshing module and copy the resulting binary in the bin folder. If you are not comfortable with cmake, make, etc. I would wait until I configure the configure.sh script properly.
pierotofy
on 6 Apr 2017
How is the cropping? Any areas that are over-cropped when compared to a ortho generated via 3d mesh?
dakotabenjamin
on 6 Apr 2017
It depends on the outlier removal parameter. The more outliers are removed, the more conservative the crop is.
pierotofy
on 6 Apr 2017
Ok, I'm familiar with make and cmake but use it with docker under archlinux. I will wait to test.
Control of outlier is planned ?
kikislater
on 7 Apr 2017
Yes, I'm adding a outlier removal parameter that can be tuned as well as a max vertex count parameter that defines the precision of the output mesh.
pierotofy
on 7 Apr 2017
Old:
2.5D:
pierotofy
on 7 Apr 2017
Old:
2.5D:
pierotofy
on 7 Apr 2017
Look nice!
Any idea why struture between 2 trees is wavy with delaunay ?
kikislater
on 7 Apr 2017
I'm not entirely sure; I'm not very familiar with the mvs texturing package, which is the component that overlays the texture over the surface model.
pierotofy
on 7 Apr 2017
Probably because of noise:
pierotofy
on 7 Apr 2017
Looks like more a shiny surface with different shine on each picture.
kikislater
on 7 Apr 2017
No that's because illumination is turned on.
pierotofy
on 7 Apr 2017
Re-ran it with half the number of vertices (less noise, but less definition on edges):
pierotofy
on 7 Apr 2017
@kikislater you should be able to test the 2.5D mesh feature by checking out the changes in #542 and running . configure.sh.
pierotofy
on 7 Apr 2017
Related issues
wolkstein
路
5Comments
ichfred
路
4Comments
nyinyinyanlin
路
3Comments
pierotofy
路
4Comments
smathermather
路
5Comments
Most helpful comment
Starting to look a lot better: