Suggested workflow for fisheye lens?

[urq...@gmail.com]

  Let me poll the field and see if anyone has a good suggestion.  One of the cameras I would like to use for photogrammetry is a Bebop 2 drone that mounts a 180 degree (but cropped) fisheye camera.  As far as I can tell, COLMAP makes an excellent sparse point cloud using the “simple_radial_fisheye” camera model.  However, the only exporter I can make sense of is for “*.nmv” which says it does not support the fisheye camera model.

 

  So far, the best process I have found is to have COLMAP run through to a complete sparse mesh, then run the “undistort” tool with min_scale set to 0.8 to capture a fairly large amount of the data available from the camera. Next, close and reopen COLMAP and run a new project using the undistorted images and the “simple_radial” camera model, then export as *.nmv.  It seems a bit of a waste to throw away all of that data, but it does work and can give some pretty good results.  At the moment I am using OpenMVS for dense reconstruction, triangulation, and texturing.  I believe it can support fisheye lenses, but I have not found a way to test that yet because *.nmv is the only format I have found that COLMAP will export that OpenMVS will import.  I am open to other tools if needed.

 

  So back to the question, can anyone suggest a better path?  I am open to ideas.

Thank you all,

Brian

[Johannes Schönberger]

Hi,

 

What is the motiviation to use OpenMVS? In my experience, the dense point cloud of COLMAP should be significantly better than what OpenMVS produces. The meshing and texture mapping is better in OpenMVS, so you can just use OpenMVS after the dense stereo stage in COLMAP.

 

Cheers,

Johannes

--
You received this message because you are subscribed to the Google Groups "COLMAP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to colmap+un...@googlegroups.com.
To post to this group, send email to col...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/colmap/c399033e-8d5f-47fb-aef9-b11b76ca5bff%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[urq...@gmail.com]

I will give it a try, since the dense reconstruction is CUDA only, I had not looked at it.  I only have access to one nVidia card, and that's an old K2200.  However, I won't be able to test that path until Monday.

 

To answer the motivation question, in general I try to avoid vendor-lock, and CUDA is nVidia only.  The other motivation is the number of machines I have access to with 64 or 128GB of ram, but no video card at all.  Slow processing is not a problem for me, but the desired projects can be huge.  I will let everyone know how it goes.

Thank you for the reply,

Brian 

[Peter Falkingham]

Hi Johannes,

  I've generally found COLMAP's dense stereo stage to be particularly demanding (in terms of hardware) and to take quite a long time, so have generally used openMVS as Brian does.  Should that be the case?  Obviously settings will matter, but I've not found a good way of reducing processing time and maintaining high quality of reconstruction.  What's generally the bottleneck for dense reconstruction?

P

[Johannes Schönberger]

Hi,

 

Have you looked at https://colmap.github.io/faq.html#speedup-dense-reconstruction ?

 

Cheers,

Johannes

 

From: 'Peter Falkingham' via COLMAP <col...@googlegroups.com>

Reply-To: <col...@googlegroups.com>
Date: Friday, April 20, 2018 at 1:59 PM
To: COLMAP <col...@googlegroups.com>

--

You received this message because you are subscribed to the Google Groups "COLMAP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to

colmap+un...@googlegroups.com.

To post to this group, send email to

col...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/colmap/f1fa9f5d-360c-4f2b-810a-19465a2485cb%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[urq...@gmail.com]

Hahaha, oh wow. Dr. Falkingham, you do realize that it is due to the research and advice on YOUR webpage that I tried COLMAP and OpenMVS to start with? It has also been an inspiration for some of what I have done with them in my spare time. (http://brian.harrison.org/Photogrammetry/JemezHistoricalSite/BodyJemez.html)

Thank you both for your work. I will follow up on the results next week.

[Peter Falkingham]

Glad it's been useful to you (but all credit goes to Johannes for making COLMAP).

Johannes: thanks.  I've played with those settings before and forgot about them.  I feel my attempts to speed up dense reconstruction generally reduced quality too much, but I'll give it another go.

[Johannes Schönberger]

window_step 2 should decrease quality just a little but result in 4x speed up. 

On 21 Apr 2018, at 20:10, 'Peter Falkingham' via COLMAP <col...@googlegroups.com> wrote:

Glad it's been useful to you (but all credit goes to Johannes for making COLMAP).

Johannes: thanks.  I've played with those settings before and forgot about them.  I feel my attempts to speed up dense reconstruction generally reduced quality too much, but I'll give it another go.

--
You received this message because you are subscribed to the Google Groups "COLMAP" group.
To unsubscribe from this group and stop receiving emails from it, send an email to colmap+un...@googlegroups.com.
To post to this group, send email to col...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/colmap/43bb2c9c-8edc-4b59-bec7-94d40d974672%40googlegroups.com.

[Brian Harrison]

To report back in:  Using the dense reconstruction as-is did not work very well.  The output mesh was missing walls and had poor coverage in several places.  In looking at it, it seems to be because the default settings on the Undistortion step take the 3320x4096 fisheye image and create a 819x664 image of only the very center of the source image.  That's only about 4% of the collected data.  It is very difficult to keep the object of interest in the perfect center of the image and it was often only partially in view.  Using the extra->undistortion option, I have access to other settings and using a min_scale of 0.80 seems to give usable coverage.  The resulting image is 3276x2656 and seems to use about 60% of the source data (math says 64.0%), though it gets a bit blurry at the edges.  It might be that all I need is access to the undistortion settings during dense reconstruction the way there is access to stereo, fusion, and meshing settings.

 For humor's sake, I tried allowing the dense reconstruction to create undistorted images, then copied the 0.80 images over them and ran the rest of the steps.  Unsurprisingly, this did not work.

I attached the source and two undistorted images to help explain.

Brian

[gavin...@gmail.com]

Brian, did you ever get this to work?
I run around interiors with a Ricoh Theta frequently, and it seems a waste to not use that for SFM reconstructions.

-Gavin

[Brian Harrison]

  At the moment, I am still using the two-pass method I mentioned at the start.  It is not great as, between the two runs, I will often have 15-20% of my photos rejected.  If I can manage to find some free time, I need to either figure out an export method that supports fisheye, or see if I can add a tab to COLMAP's dense reconstruction that includes the undistortion setting I need.  That said, direct use of the fisheye images seems like it would be best.  The sparse clouds COLMAP generates from them often look very good.

 When I look around, it does seem like there are ways to use these types of photos.  I need to find some more time to experiment.

https://github.com/cdcseacave/openMVS/issues/117

Brian