Newbie in Open Drone map for working

Hi everyone,

I am Carmen from Spain, Computer Vision and Machine Learning Engineer. I am currently working with drone imaginery for wind turbine inspection and looking for using your tool for image stitching … Crossing fingers :slight_smile:



I look forward to seeing your data! I love turbines and solar fields!


I am working on the same thing right now. I am using this and VisualSFM and I’m having a LOT of problems getting decent point clouds. I think most of the problems are with the way I am taking the pictures, and trying to keep the horizon out of the pictures a much as possible, but so far I’m finding that pretty much impossible. I’m trying to follow standard nadir/oblique techniques but it’s hard with a skinny object 100m in the air.
Any tips ?

1 Like

That’s a pretty difficult space. One thing that might help is finding an approach for automaticaly masking the sky (which might be easy with a blue sky, but more complicated with anything else).

As to skinny stuff in particular, there is some discussion in this thread:

The ODM settings there my serve as a starting point. Meshing is the real challenge, if you require a mesh.

1 Like

Thanks for tip! I really wish I understood more of the parameters than I do.
I am running almost 400 images right now with the min features at 10000. (started before I read your reply).
On my first run it excluded all of the nadir images. So I ran it again with just nadir images for an orthophoto and it actually missed almost all of the turbine. It captured the hub and part of one blade at it’s widest part.
I’m going back to the wind farm this week and am going to try to isolate just one blade in the 12:00 position and see how that goes. Both with nadir/oblique images as well as close-up “crawling” up and down the blade.


I’m not making a lot of headway…it’s actually getting worse.

It’s really having a hard time separating the blades from the rest of the world.

the point clouds for both runs are pretty similar
1 : image
This was with default settings

2 : image
with these settings: Options: min-num-features: 10000, rerun-from: opensfm, use-3dmesh: true, ignore-gsd: true, mesh-octree-depth: 12, camera-lens: brown

Ahh, so you are trying to mesh. Here’s the problem: the point clouds are decent, but include sky. There are three ways to avoid sky:

  1. don’t take pictures of it. As you know, this is impractical, you can’t get the bottom of the blades and do this
  2. mask it from the files. It looks like it’s a partially cloudy day, so this will be challenging
  3. use an approach like TSR for the mesh creation (as seen here:

This last approach is probably the only real option, but it’s a moderately heavy technical lift with some risks. TSR takes advantage of ray tracing to refine mesh creation to ensure that the mesh is refined to match reality, rather than a naive Poisson meshing approach. This means that this approach will be computationally intensive, and may not scale. But it’s a fantastic approach, and if it does work, would make ODM a best in class at reconstructions like this.


OMG, my brain hurts just trying to understand TSR.

Lol. Mine too.

So after consulting some wind turbine guys, there is very little, if any, diagnostic value to a model of a complete tower. However they are interested in individual blades. Most drone photography is done to look for visible defects in the blades alone. They even had a blade come apart without warning once. There was no obvious defect noticed in the previous pictures, however on closer examination they noticed a slight “dent” developing near the base of the blade.
So…My next project will be to try to do an accurate model of a single defect-free blade in the 12:00 position, then do model of another blade with some known defects and compare the two using cloud-compare or meshlab or similar to see how well defects turn up…in particular bulges or dents in the walls.
Never Give Up.!


That should be easier to reconstruct, and you can disregard the background sky and ground noise.

Keep us updated! This is a pretty interesting use case.

1 Like