Agisoft Metashape 1.7.0 pre-release

[Diego]

Do you use GPU for the processing, which model and which driver version? And can you provide the log from 1.7.0 related to the beginning of the processing operation which includes the actual depth maps generation (if it ever starts).

Hello Alexey,

Driver GPU Nvidia 417.22, version 1.7 finished in more than 3 hours versus 15 seconds of version 1.6.5

2020-12-08 11:35:47 checking for missing images... done in 0.009 sec
2020-12-08 11:35:47 Finished processing in 0.009 sec (exit code 1)
2020-12-08 11:35:47 BuildDenseCloud: quality = Lowest, depth filtering = Mild, PM version, point colors = 1
2020-12-08 11:35:48 Generating depth maps...
......
......
2020-12-08 14:47:12 452130 points extracted in 0.763 sec
2020-12-08 14:47:12 saved dense cloud in 0.054 sec
2020-12-08 14:47:12 loaded regions in 0 sec
2020-12-08 14:47:12 loaded dense cloud data in 0 sec
2020-12-08 14:47:12 joining 1 dense point clouds
2020-12-08 14:47:12 dense clouds joined in 0.04 sec
2020-12-08 14:47:12 loaded camera partition in 0.001 sec
2020-12-08 14:47:12 loaded regions in 0 sec
2020-12-08 14:47:12 Finished processing in 11484.6 sec (exit code 1)

Attached log file full,

Best regards,

Diego

[Alexey Pasumansky]

Hello Diego,

Thank you for sharing the log. It seems that the depth maps pre-processing step related to internal correcting of the images is taking long time on your system.
In the next pre-release update this stage will be ported to GPU (providing that at least one GPU device is selected) and it should work significantly faster.

[Diego]

Hello Alexey,

Something that also seems to happen is that in reality v1.7 does not take the region into account when selecting the cameras, in version 1.6.5 it selects 15 cameras from the region but in version 1.7 it selects 579 cameras, that is to say all , although in the end the dense point cloud is only from the region, it is that going from 15 seconds to 3 hours is crazy, even so without using the GPU, although the GPU is activated

Best regards,

Diego

[Mak11]

Diego,

As I noted in my first post in this thread it looks like the bounding box volume no longer affects the depth maps generation. for example the full frame is always used to generated the depth map even if only a small part of the image is present/needed in the bounding box & all cameras are used. This will obviously drastically affect compute time depending on the project but on the flip side the depth maps quality should be more precise & detailed. At least that's what I noticed.

Mak

[Diego]

Mak,

I understand that using the entire depth image, but using 579 cameras where the maximum region needs 15 is ridiculous, that's why I say that going from 15 seconds in version 1.6.5 to more than 3 hours in version 1.7 has no justification, something is already wrong in the code.

Best regards,

Diego

[eltama]

Hello
I got this error, latest build, JPG images only (no openxr):
2020-12-11 17:39:59 BuildModel: quality = Ultra high, depth filtering = Mild, PM version, reuse depth maps, source data = Depth maps, surface type = Arbitrary, face count = 90000000, volumetric masking = 0, OOC version, interpolation = Enabled, vertex colors = 1
2020-12-11 17:39:59 Generating mesh...
2020-12-11 17:39:59 Compression level: 1
2020-12-11 17:39:59 Preparing depth maps...
2020-12-11 17:39:59 3678 depth maps
...
2020-12-13 01:04:19 Devices performance:
2020-12-13 01:04:19  - 56%    done by [GPU 1]  with timings: 1801 s = 0% load hists + 47% load cams + 3% calcs
2020-12-13 01:04:19  - 42%    done by [GPU 2]  with timings: 1801 s = 0% load hists + 48% load cams + 2% calcs
2020-12-13 01:04:19  - 2%    done by [GPU 3]  with timings: 1801 s = 0% load hists + 60% load cams + 0% calcs
2020-12-13 01:04:26 finished group #19/189: 2946 cameras, 1826.85 s working (0% files opening + 59% IO + 41% unpacking + 0% queueing)
2020-12-13 01:04:27 Initing hists in leaves group #20/189: [512; 549)...
2020-12-13 01:04:38 total cubes: 108923782 cubes, 1662.05 MB, 144.828 MB/s
2020-12-13 01:04:38 Data: 2493.07 MB hists, 640.062 MB pyramid
2020-12-13 01:04:38 Found 3 GPUs in 0.001 sec (CUDA: 0 sec, OpenCL: 0.001 sec)
2020-12-13 01:04:39 [GPU 1] Using device: GeForce RTX 2060 SUPER, 34 compute units, free memory: 7174/8192 MB, compute capability 7.5
2020-12-13 01:04:39   driver/runtime CUDA: 11010/8000
2020-12-13 01:04:39   max work group size 1024
2020-12-13 01:04:39   max work item sizes [1024, 1024, 64]
2020-12-13 01:04:39 [GPU 2] Using device: GeForce RTX 2060 SUPER, 34 compute units, free memory: 7108/8192 MB, compute capability 7.5
2020-12-13 01:04:39   driver/runtime CUDA: 11010/8000
2020-12-13 01:04:39   max work group size 1024
2020-12-13 01:04:39   max work item sizes [1024, 1024, 64]
2020-12-13 01:04:39 [GPU 3] Using device: GeForce RTX 2080 SUPER, 48 compute units, free memory: 6852/8192 MB, compute capability 7.5
2020-12-13 01:04:39   driver/runtime CUDA: 11010/8000
2020-12-13 01:04:39   max work group size 1024
2020-12-13 01:04:39   max work item sizes [1024, 1024, 64]
2020-12-13 01:08:01 openexr error: Error reading pixel data from image file "openexr". Data decompression (zlib) failed.
2020-12-13 01:08:14 [GPU 1] Exception
2020-12-13 01:08:14 [GPU 3] Exception
2020-12-13 01:08:14 [GPU 2] Exception
2020-12-13 01:08:14 Peak memory used: 70.67 GB at 2020-12-11 18:49:49
2020-12-13 01:08:14 deleting all temporary files...
2020-12-13 01:09:03 done in 49.356 s
2020-12-13 01:09:03 Finished processing in 113343 sec (exit code 0)
2020-12-13 01:09:03 Error: Can't decompress depth

[Alexey Pasumansky]

Hello eltama,

Which 1.7.0 build number you are using, the OS version and what is the NVIDIA driver installed?

[eltama]

Hello Alexey
Metashape version: 11539, Os Win 10 Pro: build 19042.685, version 20H2, Nvidia driver:  Game studio 457.51.
Now I build mesh again with the same confiration but with Nvidia driver: studio driver 457.30

[BjFaeTorphins]

Alexey
The quality of dense point cloud that is generated in v1.7 and the filter by confidence tool is a major step forward for the underwater models that I work on. Top marks.
I have run into a bit of a dead end with a variation of the workflow I am using. I initially generated the DPC with no colour but calculating point confidence.  I then filtered by confidence to get rid of the noise.  Finally, I coloured the PDC. 
All is good except that I wanted to display the dense point cloud confidence with a view to getting rid of more noise but the point confidence appears to have been removed by colouring the dense point cloud. I have not been able to find a menu option to re-calculate point confidence.
Build 1.7.11539 standard edition on windows 10.

[LipcaCZ]

Hi,

I wanted to test that new laser scans meshing function, but it does not work for me. I've sucesfully imported and aligned the scans, but when I try to make a mesh from them, I just get error "Not enough cameras". The scene is 54 laser scans made with Leica RTC360, scans were registered in Cyclone Register 360 and exported as E57. There are no photos, just the scans (which is probably the main problem). Am I doing something wrong, or is in current version not possible to mesh laser scans alone, without any photos?
Thanks.

[Alexey Pasumansky]

I wanted to test that new laser scans meshing function, but it does not work for me. I've sucesfully imported and aligned the scans, but when I try to make a mesh from them, I just get error "Not enough cameras". The scene is 54 laser scans made with Leica RTC360, scans were registered in Cyclone Register 360 and exported as E57. There are no photos, just the scans (which is probably the main problem). Am I doing something wrong, or is in current version not possible to mesh laser scans alone, without any photos?

Hello LipcaCZ,

The possibility of using only laser scans for the surface (dense cloud, mesh) generation will be included in the next pre-release update.

[Alexey Pasumansky]

The quality of dense point cloud that is generated in v1.7 and the filter by confidence tool is a major step forward for the underwater models that I work on. Top marks.
I have run into a bit of a dead end with a variation of the workflow I am using. I initially generated the DPC with no colour but calculating point confidence.  I then filtered by confidence to get rid of the noise.  Finally, I coloured the PDC. 
All is good except that I wanted to display the dense point cloud confidence with a view to getting rid of more noise but the point confidence appears to have been removed by colouring the dense point cloud. I have not been able to find a menu option to re-calculate point confidence.
Build 1.7.11539 standard edition on windows 10.

Hello BjFaeTorphins,

Thank you for reporting, we will try to fix it in the next pre-release update as well.

Upd: Unfortunately, the fix has not been included to 11637 build, but is ready for the next update.

[Alexey Pasumansky]

Metashape version: 11539, Os Win 10 Pro: build 19042.685, version 20H2, Nvidia driver:  Game studio 457.51.
Now I build mesh again with the same confiration but with Nvidia driver: studio driver 457.30

Hello eltama,

The corruption of the depth map in the project could be related to the memory issues. I can remember only a few somewhat similar reports (with "Null Image" error message), that may indicate on the system stability issues rather than on the program bugs.

If the problem is reproducible, I can suggest to run multi-threaded Memtest86 for some long time.

[eltama]

Hello Alexey
Thank you for the tip, (I try Memtest86 with this clock frequencies and with a lower one) but now I can fix the problem with studio driver 457.30, I had to stop my simulation with BuildModel: quality = Ultra high, it will take more than 30 days probably due 61MB  x 3678 cameras, but with a BuildModel: quality = high is fine and done with only 3 days, congratualtion, much better detail then 1.65, still some "onion" layers over surfaces, in my simulation, I accidentally mix 25mm set with 24mm images, some of them are made with focus stacking tecnique, I do same simulation with 1.65 and post comparison.
Thanks

[Diego]

Hello Diego,
Thank you for sharing the log. It seems that the depth maps pre-processing step related to internal correcting of the images is taking long time on your system.
In the next pre-release update this stage will be ported to GPU (providing that at least one GPU device is selected) and it should work significantly faster.

Hello Alexey,

Thank you very much, excellent work, now it works very well in the latest build 11637.  For when the possibility apply the rolling shutter corrections for undistorted images export?

Best regards,

Diego