Insufficient overlap

What would be the options for overlap that seems a bit insufficient for reconstruction. Basically I have about a third of the dataset missing in the final orthophoto as there is most likely not enough overlap in that part as it goes over the hill…
Sometimes if I set --min-num-features to 30000 or 45000 I able to get that part seen in the final reconstruction sometimes not… it looks like with every run the ODM picks points a bit differently / randomly I don’t know… so sometimes it stitches sometimes not…

I have weird (to my knowledge and understanding) situation when I have 3 missions with 3 overlapped datasets… and they, when put together, all should stitch to each other for the final map
Something like this Dataset 1 > Dataset 2 > Dateset 3 where Dataset 2 is between 1 and 3
When I stitch Dataset 2 and 3 they stitch alright and I see both of them in the final reconstruction of 2 and 3…
But when I try to stitch them all together 1 > 2 > 3, then Dataset 1 and 2 get stitched, but Dataset 3 that previously was stitching to Dataset 2 is gone and not seen on the final map… I wonder why and if there is a way to put them all together since 1 and 3 can stitch to 2 separately but can’t when I try to put all of them together…

I will add more visual and technical details a bit later, but for now, what would be the way to go in order to compensate seemingly insufficient overlap in order to pull those bits more reliably to each other… ?

2 Likes

Increasing --feature-quality and --min-num-features and maybe also --matcher-neighbors --matcher-distance might help, but if the overlap is really too low… Not much can be done, I don’t think.

You may also have some luck with using ORB with 160k min-num-features, depending upon the scene.

2 Likes

I was lucky to find an older incomplete dataset of the same area, that crosses all those 3 missions course at about 60 degree and 50 metres bellow altitude… I tried to process it all together with 60000 min features and it all finally stitched together… I guess extra data brought extra features to match, filling the gaps it’s been having…

What is ORB anyway?
And how exactly --matching-neighbours and --matcer-distance could be adjusted here?

2 Likes

ORB is another feature detection/description engine like SIFT or HAHOG that runs much faster and lighter and usually performs similar to SIFT or better, depending.

You can increase the matcher-neighbors by 25% and same for the distance as an easy starting point.

2 Likes

Thanks a lot. I will definitely keep in mind your advice if I hit similar problems in the future.
One more question. What parameters do you think affect most the quality of texturing, especially textures closer to the edges as they always tend to be most affected by distortions of all sorts.

1 Like

Good flight planning. In other words, over-fly, or ensuring that the area of your site that you want great coverage of has a images that are outside of that so that it reconstructs properly up to the edge and you can crop off the messy extra later.

2 Likes

The areas and the sizes of them I fly sometimes hard to plan well from the first attempt. The quality of online maps either it Google or Mapbox or something else freely available, that usually used for those areas by most of flight planners is so poor, that you know only approximately where is you supposedly flight area is…
Sometimes it’s like playing strategy computer game. You send scouts to discover the land hidden in the dark, then you can plan further from what you discovered :smiley:
So sometimes I do just that… I go fly… I bring the results, I see what I got and decide if I need to fly more in case something isn’t covered or has too much distortion…
I thought maybe putting some extra computer power and options adjustment could do some quality improvements for edge areas. Sometimes it’s such a mission to go somewhere again :smiley:

2 Likes

You can try cranking up min-num-features and matcher-neighbors and matcher-distance like I suggested. GCPs might help a bit, too. But ultimately, overfly is cheap insurance!

2 Likes

Thanks a lot! You are the legend! :slight_smile:

2 Likes

I’ve just made this mistake too many times :rofl:

2 Likes

Does processing it with features quality or pc-quality set to ultra bring any noticeable improvement to the edge areas?

It can (mostly feature-quality), but likely site wide.

2 Likes

Can you explain exactly how this operates in the software? Is it computed from EXIF data to find distance centre to centre of images, or centre to edge, or something else?
If it is set to say 50m, does any matching get attempted for images outside that range at all - ie does the GPS uncertainty parameter have any influence, ie is it added to the 50m, or is it not taken into account in this process?

For matcher-neighbors, does it stop looking for matches when it reaches that number, even though there may be more images with some overlap?

2 Likes

Really gonna make me dig in the source code for this one, sheesh.

I’m not on these pages of the documentation yet! haha. I’m going in alphabetical order :sob:

Please give me a bit to dig/parse and make sure my understanding is 100% correct.

2 Likes

Thanks, take your time as there’s no real rush on it. It would be handy knowledge for fine tuning parameters on large image datasets.

Also wondering if no matcher parameters set, does brute force apply- check every image against every other image?

2 Likes

Agreed! I want to make sure I have these described quite clearly in the docs so folks have the tools they need to answer that type of question, so thank you for the prompt and confirmation that this is needed :slight_smile:

Yes, if you set both --matcher-distance and --matcher-neighbors to 0 it will skip prematching entirely.

This can be pretty beneficial for certain types of collections (orbit surveys of a cell tower, for instance), but with SIFT/HAHOG it will massively inflate the processing time.

2 Likes

We’re reaching some limits of SfM processing without semantics. Hopefully in the future, OpenSfM will have full baked semantic processing capatibilities which will allow :

  • Removing points in the sky, which makes fuzzy and distants points that breaks auto-crop
  • Removing vegetation points for cell tower processing
3 Likes

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.