Agisoft Metashape 1.6.0 pre-release

[Paulo]

Just an update on Metashape v1.6 processing of satellite images...

With latest release, I still find big gaps in depth maps and dense cloud on Tripoli sample image dataset....

[wojtek]

Hey Paulo,

But you only have 2 images there?

[cbnewham]

Just trying this new version.

I used a model I've previously built on High. This time I tried Ultra and it worked with 2,000,000 faces, but when I upped this to 6,000,000 faces to produce a more detailed mesh I got "Assertion Failure at line 118" after processing for a couple of hours (BTW, I selected "reuse depth maps" on that second build).

Also, this version has the out-of-core processing. I presume that is using Disk to offload things to? There don't seem to be any settings for this?

Good work - looking forward to the next release.

[Alexey Pasumansky]

I used a model I've previously built on High. This time I tried Ultra and it worked with 2,000,000 faces, but when I upped this to 6,000,000 faces to produce a more detailed mesh I got "Assertion Failure at line 118" after processing for a couple of hours (BTW, I selected "reuse depth maps" on that second build).

Also, this version has the out-of-core processing. I presume that is using Disk to offload things to? There don't seem to be any settings for this?

Hello cbnewham,

Do you have the related processing log that you can share?

The out-of-core mesh generation works for PSX projects, providing that Fine-Level Task Subdivision is enabled in the Advanced preferences tab.
Currently the temporary data is saved in the project.files folder. Also at the moment there's a tweak, that allows to define the temp folder for the mesh generation.

[Mak11]

Alexey,

No trimming tweaks were applied. If you try the Doll sample you will see similar results (holes) where there's poor depth map coverage (under the skirt) and  mesh_trimming_radius 30 doesn't seem to have any effect when applied in this build. (obviously 1.5.0 handled does cases perfectly without any tweaks...but you already know that 😉). Visibility Consistent Mesh Generation, Metashape up to v 1.5.1 ( & to a lesser degree Metashape 1.5.3+ with mesh_trimming_radius set to "30") correctly remeshed those holes/gaps. So there must be a way to fix this.


Yeah I'm seeing some GPU utilization but what appears in the console has changed quite a bit so it was harder to figure out what was happening.

Out-of-core meshing is perfectly working though. So big congrats on this achievement! Never went over 14GB of Ram usage on data-sets (with 150+/200+ 24MP images) that would fail before.

Cheers

MAK

Cheers

Alexey,

After further testing it unfortunately does look like the main/mesh_trimming_radius tweak doesn't have any effect anymore.

Example the Doll: With and Without main/mesh_trimming_radius tweak (set to 0 or 30 or 90) the mesh is exactly the same with exactly the same number of polys and vertices:



Cheers

Mak

[Alexey Pasumansky]

Hello Mak,

You can disable trimming of the out-of-core method by using main/mesh_tgv_ooc_trimming_enabled tweak and setting its value to False.

If you want to switch off out-of-core mesh generation method and use the one from the latest 1.5 release, then you need to set to False another tweak: main/mesh_tgv_ooc_enabled

[Mak11]

Alexey,

Interesting. Just tested with the doll and it did indeed close the holes (also went crazy on top but this was sort of expected given that the mesh was already poorly reconstructed.)



I'm now testing on an other project to see how it turns out.

Maybe there's a better balance to find with the out-of-core method trimming and make it less aggressive (instead of having to turn it off/False) ?

Will post my finding once they are processed.

There's no way I'm going back to non-out-of-core method 😮😊. I've been waiting for it for years and its finally delivered (and is a god send for some projects). Also the reconstructed mesh details is the best so far. Only thing left to fix/tweak is those holes/vertices issues.

UPDATE:

Looks a lot better with main/mesh_tgv_ooc_trimming_enabled "FALSE" ! And without any messy polys running inside the mesh or all over the place like in 1.5.2+.




Cheers

Mak

[cbnewham]

Hi Alexey,

I set the number of faces to 4,000,000 and it didn't  generate the error again. I will post the log if it happens again.

Thanks!

[Alexey Pasumansky]

Just an update on Metashape v1.6 processing of satellite images...

With latest release, I still find big gaps in depth maps and dense cloud on Tripoli sample image dataset....

Hello Paulo,

Looks like we were able to reproduce similar problems and will implement the fix in the next pre-release update.

[Paulo]

Great, Alexey,

Thanks for the update

Just to confirm that RPC satellite processing of Tripoli Digital Globe data sample no longer encounters gaps in dense cloud processing... Thanks for the good work!

[Mak11]

Another issue with the new out-of-core mesh generation is that there is some strange inconsistencies in the mesh topology for no apparent reason at all. It varies every time when a mesh is generated even with the same depth maps correction meant "using the same data set & generating new the depth maps":

Extremely low polycount in some places:



Regenerated the mesh and: 



MAK

[wojtek]

Is there any way to force more UV Island stitching? (at the expense of extra deformation?)

The new uv parameterization is very nice but dense meshes get broken down a bit too much.

[BjFaeTorphins]

I was wondering about the new options in the align photos stage : source - estimated - sequential.
I think sequential speaks for itself but what do the others ?

Sequential pre-selection is trying to match the images with a bunch of their neighbors according to the camera labels.

Estimated preselection is based on the estimated camera locations after initial alignment (for example, performed using rough accuracy settings).

One of the possible scenarios of use - when aligning the video sequence frames run Sequential preselection at first and then Estimated preselection, if there are intersecting parts of the image acquisition path.

I've given this a try. 

Sequenial Preselection certainly is much quicker. 6 minutes for 594 cameras compared with 81 minutes for "Estimated".  Both runs reset alignment.

Running the alignment a second time with Estimated after first doing a Sequential didn't seem to do anything.  (Both done at High accuracy)


Is the idea that the Estimated (second alignment) should improve the accuracy?

Can you explain the steps?

[Alexey Pasumansky]

Hello BjFaeTorphins,

Using Estimated preselection option after quick Sequential alignment run may be reasonable, for example, if the frames are acquired from the vertical zig-zag survey. Sequential preselection option would not give you tie points for the images from the different flight lines, therefore further processing may result in the artifacts (like stairs) on the surface in the areas of the overlap between the flight lines.

[BjFaeTorphins]

Hello BjFaeTorphins,

Using Estimated preselection option after quick Sequential alignment run may be reasonable, for example, if the frames are acquired from the vertical zig-zag survey. Sequential preselection option would not give you tie points for the images from the different flight lines, therefore further processing may result in the artifacts (like stairs) on the surface in the areas of the overlap between the flight lines.

After alignment with Sequence when I try  a second alignment using Estimated I see the following message...
"Warning: Can't resume matching without keypoints"
What am I doing wrong?