Open Source Photogrammetry with Meshroom

meshroom
photogrammetry

(Nebulousflynn) #1

Getting some decent results with Meshroom (https://alicevision.github.io/#meshroom) - 80 images processed on default settings:

A 100K crop of the fountain (the intended subject of the scan), 8+ hours processing:

To compare with a RealityCapture version, less than an hour or so processing:

The node based workflow feels pretty powerful… but I have no idea what a lot of the parameters do :slight_smile:

would love to see (and maybe I just haven’t looked hard enough):

  • a cropping step/tool after alignment and before depth mapping
  • a pause button
  • basic editing tools

(Beguinraphael) #2

Nice discovery here, but at this time thre’s no doc n’or just an quick howto on her youtube.

SOrry if it’s not the place, but how you’v got your avatar picture?

Txs in advance :wink:


(Tom Goskar) #3

Promising results and great to give a comparison. Must give it a go too.

What are you using to rebake these days?


(Mesheritage) #4

For the parameters it is very similar to openMVG for the SFM process, from what I understood it comes from the same project initially.

Those who are using openMVG should fell very familiar, but just as a note, you need to have CUDA for the dense & Co (a bit like colmap).

One thing that can be done to speed thing up is to change the matching method. ANNL is very slow compare to FAST_CASCADE. ANL works better but from my experience, it is worth using it only when you have bad dataset.
For example, this model needed ANNL (fix camera, lack of light, video,turning hand, with strong background…Not ideal!)

But even with openMVG I mostly use FAST (default)


(Nebulousflynn) #5

@beguinraphael the project is still young and I imagine documentation will grow as people get to grips with it. Sketchfab user Peter Falkingham wrote a post sharing his experience testing the software: https://pfalkingham.wordpress.com/2018/08/11/photogrammery-testing-14-alicevision-meshroom

We hope to share a tutorial on our blog soon too. To be honest I just threw in the images and hit ‘Start’ though!

@tomgoskar generally I use Blender to decimate/retopologise + bake textures. If I use two meshes straight from photogrammetry software (i.e. high and low res exports) then I’ll just use xNormal to create normal and ambient occlusion maps for the low res model.

@Mesheritage thanks for the info! I’ll have to try changing the matching method.


(Nebulousflynn) #6

I’m going to keep posting examples of models made with Meshroom here as I find them


(Beguinraphael) #7

understood thanks for your details and the blog !


(Nebulousflynn) #8

(Alensdarkly) #9

Love the look of this!

Is there a version for MAC IOS 10.13?

Thank you


(The Net Striker) #10

I’ve also tried this software and for static objects it is working very good. I’ve also tried to scan myself by just taking a few photos from multiple angles while i was turning. (just like i was standing on a turntable) But that didn’t work for some reason.

The photos that were taken from the face did light up green, but all photos from the back of my head did light up red. And the face was also not recognizable in the final model. I’ve also tried to remove the background and save the images as png, but that didn’t help either.

Is there maybe a setting to tell the software that this is like a turntable scan?


(Nebulousflynn) #11

a few more examples!


(Zombie Farm) #12

Wow… looks like people are getting some great results. Does anyone know of a way to have it only calculate a defined areal (like in Reality Capture?)… ie I don’t need the whole room and would like to be able to bump up the resolution on just what I need.


(Nebulousflynn) #13

I’m bugging the dev team at Meshroom for a tutorial and hope that they include this - I had the same thought while I waited 8 hours for a big scene to compute when I really only wanted a small part of it…


(Michal Zurawski) #14

I see you use Meshroom extensively. Has it any photogrammetric features? Title of thread is pointing that yes, it has. Also official website is telling “Meshroom allows you to run the whole photogrammetric pipeline”… at the same time it mention that this is SFM. So what is the true? Photogrammetry or SFM?


(Steren) #15

I found this issue that discusses documentation, and it actually links to a google doc with 90+ pages of documentation (that is very hard to digest)

I am really hoping that Meshroom maintainers would invest more in community management and documentation. Today, I find it impossible to use Meshroom outside of the default settings.


(Sim) #16

Hi @steren, I´d love to get some more feedback on the documentation (just leave comments in the doc).
A few video-tutorials are also on my to-do list.


(Steren) #17

Regarding the doc: the better is the enemy of the good. My main feedback is: migrate it as soon as possible to gitHub and host is on a web page, so that it can be indexed by search engines and so that others can send pull requests.


(Nebulousflynn) #18

I have no idea about what algorithms / processes exactly power Meshroom.

If you wish to make 3D models from a digital image set then you can do so with this software.


(Sim) #19

@michal_zurawski and @nebulousflynn A general overview of the photogrammetry pipeline can be found here. Meshroom is a SfM Photogrammetry software.
You could differentiate traditional (analog) photogrammetry and SfM photogrammetry. SfM photogrammetry uses the same approach but utilizes automation and new algorithms. For details read this paper (p.1-2).


(Michal Zurawski) #20

So how do you measure an object in Meshroon. How you can imput an reference and points to confirm that aligment is correct. What about an error of reconstruction? Because if, the software “photogrammetric” it has to allow use of these kind of features. And you are using hybrid “SFM photogrammetry”… what it means?

http://www.isprs.org/tc2-symposium2018/images/ISPRS-Interview-Fraser.pdf

“The names of SLAM, SfM and photogrammetry do mean different things… ‘SfM’ may sound sexy, but it ain’t photogrammetry!”

And in paper that you have linked, they use also coded referenced target to make measurement and then they compare it with SFM to check.

I’m asking about possibility to use targets, scale bars etc. For an exapmle, can you mark a point on one photo and check if this point is automathicly placed on the rest of the photos correctly?