When I processed imagery in May 2019 with the WebODM version then (?), my dataset containing small overlap (55%) and sidelap (30%) images, stitched an orthophoto of the complete area, successfully!
When I retried recent images and the images of 2019 with the latest WebODM version, orthos only stitched partially.
Which parameters may need to be changed to produce legacy results?
Wow, so much has changed since then, it’s hard to say.
Are you able to post your data on someplace like dronedb.app (or another Cloud sharing provider)?
I’d start by maybe tweaking the matcher-distance and matcher-neighbors, maybe turning them off entirely.
Also, try disabling resize of images, increase feature-quality, and maybe bump min-num-neighbors a bit?
Thanks for taking the time to look into this.
It would be valuable to know if any default params. changed since mid-2019 and focus on those but if you can play around with this focusing on your suggestions, I would appreciate. I don’t have time immediately to do this troubleshooting.
Herewith a dataset that is crucial to get stitched: 181 nadir images over quite undulating terrain.
Images will be finished uploading to the folder ~12:00 PM today (CAT/UTC+02:00; 23/06/2021)
Options: crop: 0, debug: true, feature-type: hahog, matcher-distance: 0, matcher-neighbors: 0, min-num-features: 32000, pc-quality: high, verbose: true
report.pdf (5.4 MB)
You’re getting 26 Reconstruction Groups, so the processing pipeline is really chunking your data up into small clusters of about 7 images each on average. Ideally, you’d only have one or two Reconstruction Groups.
I’ve done a number of iterations and I am not getting much more to stitch than a handful of images in this region. Unfortunately, I think the overlap/sidelap might be too low in this dataset to be reliably reconstructed.
I’m even unable to get a full stich from Microsoft Image Composite Editor, which normally is about as un-fussy as you can get with input data for a naive 2D stitching software.
Thanks a lot. This is better than the smaller patches I got with default params.
I will now continue to experiment with the params. and the other datasets I have, that were stitched completely in 2019. This terrain is very undulating thus, I am not surprised by the partial output.
Kind regards.
Would appreciate if others can also provide some input.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.
