Ortho is partially warped and missing most of the original imagery

We have a series of approx 2500 images from a manned flight and are using ODM to generate an orthophoto. I’ve been told that we believe the resolution was around 3-5 cm/pixel, flight altitude was roughly 1200 feet AGL, flight time was about 1 hour, and flight area was approximately 3.5 square miles.

The only option specified on the command line was --fast-orthophoto. After failing on AWS instances with up to 160G RAM, it finally finished in just under 20 hours using an instance having 244GB of RAM.

The resulting ortho seems to include approx 20-30% of the entire area of images that was taken and I’m hoping that by providing the full log file and resulting image, someone may have some advice. If I can provide any other files or information to help, please let me know.

ODM Log File

This next image shows shows a POINT where each image was taken (left), an image from output of ODM in MeshLab (middle) and the resulting ortho (right).

PNG

I have a smaller dataset (having 900 images) for a very similar type of area which I’m using to test min-num-features of 12000 to see if I get a better result.

We are also going to start testing with ODM clustering to split projects.

I’ve seen this problem before, though in my case it was me testing what would happen by putting two surveys at different altitudes together (33m vs 100m). It resulted in only the images at 33m being stitched while the rest were dumped. It’s something that I’m going to do more testing on when I get chance:

https://dev.aerosurvey.co.nz/public.cgi?S=342&T=e1818b6c-1466-4585-9d87-0626bc1b8365

Were all of your photos at the same altitude? What was the overlap?

I increased the min-num-features count to 16000 and the mesh-size to 1000000 but it didn’t help. Maybe they need to be even higher. Increasing either of those will likely increase the RAM requirement so you’ll probably need a bigger node. I’m keen to see if increasing to 12000 works for you though as it would answer some questions that I have about large datasets…

Thanks for your time and answer! I read a lot before posting, but the result didn’t seem to fit the case where more features would help.

These images were taken from a manned aircraft and they were all taken at the same AGL - as well as the pilot could keep it steady. I did see that ODM reported the images had altitude data.

Since it takes so long to run this dataset, my next step will be to reduce the time to test. Using lat/lng’s I will split the images into 8 groups and split them using a cluster of 8 nodes. I will report the results - let’s see if we will get more of the imagery to appear.

If you see the point cloud in full, the reconstruction has completed correctly and increasing min-num-features will not help.

It seems like it’s a problem with the texturing. How does your textured 3D model look like?

It’s possible that the Z-axis is inverted (the model might be upside down, that’s why the orthophoto does not render completely).

This was run using fast-orthophoto so it generated odm_texturing_25d. I have these files in the folder - how can I open the 25d file?

capture

I’m also re-running it now without using --fast-orthophoto so I should get 3d model right?

Have you tried opening with meshlab ?

1 Like

Carloz, thanks!! I was finally able to figure out to use meshlab!

1 Like