Multi GPU vs Multi server

[Shepah]

Hi,

Currently i'm using Photoscan with big GPU rigs (Titan, TitanXP, K80...) and I see that the "filtering depth maps" is one of the longest step we have.

I wonder if any of you know what's better in performance : a big multi GPU single rig or a cluster of small servers with GPU.

For example one of our rig is 5xtitan GPU with 44 cores CPU:

Would it be better to leave it as one or make a cluster of 5 servers with 1x titan each  in term of time of processing ?

Thanks in advance for reading.

[Alexey Pasumansky]

Hello Shepah,

GPU is only used for the depth maps calculation on that stage, whereas filtering is performed on CPU only.

Long filtering time may be related to the excessive overlap between the images. Can you specify what is the type of your project - is it common nadir aerial survey, oblique aerial shooting or close-range project?

[Shepah]

Hi Alexey,

We currently use it for full body scanning, we use a rig of 110 cameras to take pictures.

I thought the more overlapping, the best it was for photoscan as it was easier to "reconstruct" the model.

[Alexey Pasumansky]

Hello Shepah,

Then it shouldn't be very long. Do you have the processing log with the dense cloud generation stage timing?

[Shepah]

Hello,

Here's a first log of activity, i'll get detail with another run after that:

Code: [Select]

16:38:16: Creating new document...
16:38:16: Adding chunk...
16:38:16: Loaded 110 images.
16:38:16: Matching photos...
16:39:45: Aligning cameras...
16:40:00: Building points...
16:40:01: Optimize cameras...
16:40:02: Building points again...
16:40:02: Reloading photos (without pattern)...
16:40:05: Detecting markers...
16:41:44: Reloading photos (with pattern)...
16:41:47: Loading markers positions...
16:41:47: Updating scene transform...
16:41:47: Updating region [1]...
16:41:47: Building dense cloud...
16:49:35: Updating region [2]...
16:49:35: Building 3D model...
16:50:16: Removing flying junk...
16:50:17: Closing holes...
16:50:17: Fixing topology...
16:50:18: Smoothing model...
16:50:18: Decimating model...
16:50:23: Closing holes...
16:50:23: Fixing topology...
16:50:23: Building UV coords...
16:50:28: Reloading photos (without pattern)...
16:50:31: Disabling part of cameras for texture...
16:50:31: Building texture...
16:51:07: Exporting model...
16:51:08: Saving document...
16:51:11: Whole process ended successfully. Closing Photoscan.

Actually in the process we use 220 photos: one with normal light, another with textured light (pattern) for better mesh reconstruction

[Shepah]

Here's a detailled log of photoscan process:

https://pastebin.com/cSQGCVAT

[Alexey Pasumansky]

Hello Shepah,

According to log dense cloud generation took about three minutes (1 for depth maps and 1.5 for depth filtering). Whereas mesh generation takes about 10 minutes.

Depth filtering and mesh generation currently are utilizing CPU only.

[Shepah]

Thanks for your quick analysis !

Is there possibility to reduce this global time more ?

The next update of photoscan will be on mesh generation using GPU ?