Agisoft Metashape 1.7.0 pre-release

[Paulo]

I have seen that the depth maps generation has a complete new  algorithm in new version. However I have tested on a set of 280 UAV images and it seems that process is much longer than 1.6.5...

It reached 7% progress and estimated time to finish was at 13 hours and going up so I cancelled and reinstalled v.1.6.5 to complete depth maps in :

Depth Maps
Count 280
Depth maps generation parameters
Quality High
Filtering mode Aggressive
Processing time 4 hours 29 minutes
Memory usage 1.70 GB
Software version 1.6.5.11249
File size 1.37 GB

so new approach may be interesting but if it takes such more time than it is not feasible...

My system is:

Software name Agisoft Metashape Professional
Software version 1.6.5 build 11249
OS Windows 64 bit
RAM 15.95 GB
CPU Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz
GPU(s) GeForce GTX 670MX

[dpitman]

Paul, thanks for sharing.  Efficient processing is a huge factor as image sets get larger and larger.  Simactive's Correlator3D is making good progress on reducing reconstruction processing time of large images sets, so it is possible.  Perhaps Agisoft has some optimization coming as well.

[jedfrechette]

[For laser scans] Current approach is the following:
...
- run Align Photos operation as usual.

Does that mean that the poses of the laser scans will be adjusted in addition to the internal and external parameters calculated for the standard cameras?

As a user who commonly combines laser scan data with photogrammetry that is not at all what I would want to happen. The laser scan data will almost always be coming from other software where we have much more control over the alignment process. Therefore the laser scan data should be treated as fixed (no pose adjustments) and photos aligned to it.

I haven't had a chance to test the implementation in Metashape 1.7.0 extensively yet. However, my past experience with the general method of treating scans as spherical panoramas for alignment purposes suggests that the results will likely be very unstable. The resulting panoramas are typically low resolution and, if they are intensity only scans, relatively low contrast; so few tie points are likely to be detected by Metashape's feature detector. From the few small tests I've tried on 1.7.0 that is indeed what seems to be happening.

From a process efficiency perspective, what I want to do as a user is use intensity only scans to capture geometry and photogrammetry to capture color.

[iidancom]

I am comparing 1.7 with 1.6.5. 1.7 is much slower (54 minutes vs 40, in my case of creating a 3D model (OBJ).
It uses also much more memory
Depth map creation 3.5 GB and 20 minutes  vs 1 GB and 10 minutes
Orthomosaic Reconstruction  9 minutes and 6.8 GB vs 3 minutes 30 seconds and 2.7 GB
GTX 1051 TI, 32GB

[laurent.G]

Hello ,
I ve try to install pre-release on RYZEN 3900x Windows 10 but installer finish without install metashape.
What can i do ?
Thank you

[dcobra]

I can't seem to use GPU for depth map generation.  I get the following error:

Warning: Device GeForce GTX 1070 Ti ignored because it has 5958/8192 MB available, but 6870 MB required

Log up to that point is attached.  I have two identical GPUs and haven't had any issues like this in the past. 

[ferengi]

hi,

occlusion map generation takes for ever

2020-10-27 13:37:53 postprocessing atlas... done in 5.486 sec
2020-10-27 13:38:02 Finished processing in 200.162 sec (exit code 1)
2020-10-27 13:38:51 BuildTexture: mapping mode = Keep uv, page count = 1, texture size = 8192, texture type = Normal map, source data = Model, transfer_texture = 0
2020-10-27 13:38:51 Generating textures...
2020-10-27 13:38:51 Initializing...
2020-10-27 13:40:10 Baking textures...
2020-10-27 13:40:10 Computing tangent vectors...
2020-10-27 13:40:11 Computing normals for baking...
2020-10-27 13:40:11 Transferring attributes...
2020-10-27 13:40:23 Applying textures...
2020-10-27 13:40:32 Finished processing in 101.502 sec (exit code 1)
2020-10-27 13:45:29 BuildTexture: mapping mode = Keep uv, page count = 1, texture size = 8192, texture type = Occlusion map, source data = Model, transfer_texture = 0
2020-10-27 13:45:29 Generating textures...
2020-10-27 13:45:29 Initializing...
2020-10-27 13:46:36 Calculating ambient occlusion...
2020-10-27 15:04:31 Finished processing in 4742.23 sec (exit code 0)
2020-10-27 15:04:31 Error: Aborted by user

[RHenriques]

Cuda is not active in MacOS. Nvidia GPU's appear as OpenGL. In my case I have two and only one appears available in the GPU list.

Hello Renato,

Do you have cuda-enable tweak set up via Advanced preferences tab?

Please also provide the output of the following command executed via the Console pane and specify the CUDA driver version installed:

Code: [Select]

Metashape.app.enumGPUDevices()

Hi Alexey

Yes, I have the tweak activated for Cuda. I've included the outputs from "Metashape.app.enumGPUDevices()" from 1.7 and 1.6.5 for comparition.
There's clearly a difference between the two versions in number of GPU's detection.
Best Regards

Renato

 

[Darko]

  Hi Dmitry
I've run into the problem with the Metashape 1.7.0. Build 11340 - WIN/PRO version. On some projects it makes the Tiled model while on some it exits immediately with the message :
Assertion 23910910010 failed at line 54
It seems that it appeared after changing the Region properties like resizing, moving etc.
What is the solution?
Best, Darko

Should I mention that the Tiled Model was started right after the Alignment process with the Depth Maps option set to Ultra High and Faces to Low while the pixel size was left as it was.

[Alexey Pasumansky]

I have seen that the depth maps generation has a complete new  algorithm in new version. However I have tested on a set of 280 UAV images and it seems that process is much longer than 1.6.5...

Hello Paul,

It would be helpful to see the processing log related to the depth maps generation in the version 1.7 pre-release.

[Alexey Pasumansky]

[For laser scans] Current approach is the following:
...
- run Align Photos operation as usual.

Does that mean that the poses of the laser scans will be adjusted in addition to the internal and external parameters calculated for the standard cameras?

As a user who commonly combines laser scan data with photogrammetry that is not at all what I would want to happen. The laser scan data will almost always be coming from other software where we have much more control over the alignment process. Therefore the laser scan data should be treated as fixed (no pose adjustments) and photos aligned to it.

I haven't had a chance to test the implementation in Metashape 1.7.0 extensively yet. However, my past experience with the general method of treating scans as spherical panoramas for alignment purposes suggests that the results will likely be very unstable. The resulting panoramas are typically low resolution and, if they are intensity only scans, relatively low contrast; so few tie points are likely to be detected by Metashape's feature detector. From the few small tests I've tried on 1.7.0 that is indeed what seems to be happening.

From a process efficiency perspective, what I want to do as a user is use intensity only scans to capture geometry and photogrammetry to capture color.

Hello Jed,

Yes, it can be said that the laser scans are aligned together with the common images.

If you need to rely on the accurate scanner locations, you can input corresponding coordinates to the Reference pane for the related camera instances and adjust the accuracy value for them in order to increase the weight for them.

And you are right that the intensity maps would provide lower number of tie points compared to spherical panoramas, however, manually placed markers (on the scanner data and on the digital camera images) could assist the alignment process, if there automatically detected tie points are not sufficient.

[Alexey Pasumansky]

I am comparing 1.7 with 1.6.5. 1.7 is much slower (54 minutes vs 40, in my case of creating a 3D model (OBJ).
It uses also much more memory
Depth map creation 3.5 GB and 20 minutes  vs 1 GB and 10 minutes
Orthomosaic Reconstruction  9 minutes and 6.8 GB vs 3 minutes 30 seconds and 2.7 GB
GTX 1051 TI, 32GB

Hello iidancom,

New depth maps generation process will be optimized for the final 1.7.0 release and actually considerable performance improvement is expected in the next pre-release update.

As for the orthomosaic generation time difference, I assume that it is mainly related to the ghosting filter option used in 1.7, which was not available in the version 1.6. Have you compared the timing with the disabled ghosting filter?

[Alexey Pasumansky]

I ve try to install pre-release on RYZEN 3900x Windows 10 but installer finish without install metashape.
What can i do ?

Hello Laurent,

Are you observing any errors during the installation? If there are any, please provide the related screenshots or specify exact error message observed.

[Alexey Pasumansky]

I can't seem to use GPU for depth map generation.  I get the following error:

Warning: Device GeForce GTX 1070 Ti ignored because it has 5958/8192 MB available, but 6870 MB required

Log up to that point is attached.  I have two identical GPUs and haven't had any issues like this in the past.

Hello dcobra,

Please check, if it helps to set up the following teak to 40 value, via Advanced Preferences tab -> Tweaks dialog:
main/depth_pm_max_neighbors

[Alexey Pasumansky]

occlusion map generation takes for ever
2020-10-27 13:45:29 BuildTexture: mapping mode = Keep uv, page count = 1, texture size = 8192, texture type = Occlusion map, source data = Model, transfer_texture = 0
2020-10-27 13:45:29 Generating textures...
2020-10-27 13:45:29 Initializing...
2020-10-27 13:46:36 Calculating ambient occlusion...
2020-10-27 15:04:31 Finished processing in 4742.23 sec (exit code 0)
2020-10-27 15:04:31 Error: Aborted by user

Hello ferengi,

How many polygons are there in the source and target models used for the occlusion map generation? Also please specify the parameters used for the mesh reconstruction for both models (if they are different) - source option, method used and interpolation.