I performed a lot of tests in the past with the available free or open-source photogrammetry solutions available.
And yesterday I tested the new Regard3D version so I think it's time to show you these "never published" tests.
Here is the list of the software that I have used for my tests (have a look at my previous post for the links) :
- MVE (Windows version)
For all these tests I have used the same dataset (photos) and ran all the programs with their default settings.
This dataset consist of 10 photos that I shot 4 years ago with a 8Mb pixels camera (Nikon Coolpix L16).
I selected this dataset because the photos are good for this study :
The subject is a bas-relief of the bulgarian poet Christo Botev that can be seen outdoor at the Bulgarian Cultural Institute in Paris where I have studied the bulgarian language and culture during 3 years (thanks to my teacher for her patience and professionalism !). A small keychain printable version of it is available for download on my Thingiverse account.
- in focus
- without contrast (sun or shadows)
- easy subject (nearly all the parts can be seen)
- not too many photos (neither too few)
This is not a tutorial on how to use the different softwares so I won't explain that part (and if you wish to know how, there's a lot of already very good explanations over the web).
It's just here to show you the different results you can expect, along with my own opinion on ease of use and the result's overall quality.
I start with that one because it's the first photogrammetry software that I tested back 4 or 5 years ago.
VisualSFM is quiet straight forward to use (it's globally a 4 steps process) and along with CMVS it will produce a dense point cloud (which will have to be meshed if you want a 3D polygon model).
In order to use OpenMVG you will need a Unix OS. OpenMVG will produce a sparse point cloud from your photos, and from there MVE will be used to produce a dense point cloud and polygon mesh.
You can either run MVE on your Unix machine or on Windows (I only used the Windows version for these tests).
This pipeline is not straight forward and might discourage those that don't want to deal with Unix commands or solve the different problems that could occur (I, for example, wasn't able to use it directly because my camera wasn't in the database. Well it was indeed... but for some strange reason my Exif data differed from the one in the database because my camera model's name was in upper case in my Exif data, and in lower case characters in the OpenMVG's database). But if you can deal with these small problems, the OpenMVG+MVE pipeline will produce a very nice mesh !
For those who have been discouraged at using the OpenMVG route above, this might be a much easier solution.
Once again this pipeline is not straight forward but at least you won't need Unix as it runs on Windows. But you will still have to launch a serie of Windows shell commands in a certain order to get to the result (this could of course be automated). Once again the result will be a very decent 3D mesh !
This is one of the easiest pipeline. All you have to do is to copy your photos in a folder and launch a batch (run.bat) by a simple double-click on it. So it's something that anybody can do, and the resulting mesh might be good enough (and without holes except the biggest one at the back) but of course if your dataset was also good (GIGO rule).
The latest version has been made available only 2 days ago, and from my last test it's able to produce a very good 3D mesh.
This program will probably seduce most of the users because it offers a very nice GUI (graphic interface) that let you see your results right away. It also saves your project in its own format which is convenient if you want to modify it later. And at least but not last, it let you try diferent algorithms and you can always come back on the previous steps to try something else (everything being stored in the project). This approach is very user-friendly and very similar to the Photoscan approach (from Agisoft).
using CMVS reconstruction (no holes)
using MVE reconstruction
I tested MicMac quiet a long time ago on this same dataset (and others) but unfortunately at that time I didn't take note of the time it took to compute the sparse and dense point clouds, neither the different options that were used. But I also upload the results here so you can visualy compare.
MicMac is a photogrammetry suite from the IGN which offers a LOT of different tools therefore offering the user a LOT of options. It produces a sparse and a dense point cloud.
The remeshed version (shown in the 2nd image) is the result of a Delaunay triangulation of MicMac's dense point cloud (shown in the 1st image) performed in CloudCompare.
As you can see on this last picture the different solutions also differ on the final mesh's orientation and scale.