Forum

Author Topic: Processing times version 1.5.2 vs 1.4  (Read 2352 times)

Paulo

  • Hero Member
  • *****
  • Posts: 1303
    • View Profile
Processing times version 1.5.2 vs 1.4
« on: March 24, 2019, 01:22:05 PM »
Hello all,

i am comparing processing a full body model using 120 photos in versions 1.5.2 and previous 1.4.5..... on same computer

2019-03-24 02:56:47 Platform: Windows
2019-03-24 02:56:47 CPU: Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz (laptop)
2019-03-24 02:56:47 CPU family: 6 model: 58 signature: 306A9h
2019-03-24 02:56:47 RAM: 16.0 GB
2019-03-24 02:56:49 OpenGL Vendor: NVIDIA Corporation
2019-03-24 02:56:49 OpenGL Renderer: GeForce GTX 670MX/PCIe/SSE2
2019-03-24 02:56:49 OpenGL Version: 4.6.0 NVIDIA 417.01

I am generating a mesh from depth maps in high quality with strict volumetric masks on….

For version 1.4.5 the time to generate  mesh from depth maps is 49 min.

For version 1.5.2 the time to generate  mesh from depth maps  is  1 h 42 min.

I realize the mesh from depth maps algorithms have changed but the considerable increase in processing time is concerning….

Hoping to get your thoughts on this issue….

Attached are processing logs from both versions and screen shot of both models side by side….

Seems model 1.52 is a little better but substancial increase in processing time does not seem to justify such marginal quality improvements….

Note that in both cases following tweak was used in preferences/advanced :  main/dense_cloud_max_neighbors  60
« Last Edit: March 24, 2019, 02:01:03 PM by Paulo »
Best Regards,
Paul Pelletier,
Surveyor

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14813
    • View Profile
Re: Processing times version 1.5.2 vs 1.4
« Reply #1 on: March 25, 2019, 09:55:00 PM »
Hello Paul,

Old experimental meshing method could work faster on the small datasets, but it had poor scalability depending on the number of cameras and was relying on a single CPU core for the main processing step.

New method gives major speed up on high end GPU and on multi-GPU systems. Also behaves well on large datasets.

I've tried to run the new meshing method on desktop with single GTX 980 Ti (the alignment might be a bit different, but other settings were the same) and got 8.5 minutes for depth maps and 6.5 minutes for new mesh (with strict masks).
Best regards,
Alexey Pasumansky,
Agisoft LLC