How much of GPU should be engaged?

[kirk]

I see only 15-20% of my GF 980ti 6gb  while doing dense cloud "reconstructing depth" process.    Is it normal?   I thought it should take advantage of GPU, right?

[Alexey Pasumansky]

Hello kirk,

Which application you are using to monitor the GPU activity?

[kirk]

I tried it with current Metashape 1_5_0_64x  and then went back to Photoscan 1_4_4_64x.

Those 15-20% is from 1.4.4  and with current Metashape 1.5.0 I have even less GPU loading  , actually hardly any load at all , maybe 3-4% during dense cloud building.

Metshape does it just forever    The GPU checkbox setting was checked in in both

Should I try to update gpu driver?  It's 416.34 currently, not especially old

[Alexey Pasumansky]

Hello kirk,

I have meant, which utility you are using to track the activity of the GPU?

In case it is Windows 10 Task Monitor, try to expand the GPU activity page and switch one of the graphs to Compute_0 or Compute_1.

[kirk]

Here it is

 I did install current nvidia driver.   Again it's a dense cloud generation.    Wonder if GPU acceleration  really works as it should with my  "Compute_0" around 50%?   Isn't it supposed to be closer to your example with nearly 100% loading or it's something  Quadro specific?     

Can I do something to squeeze more of my GPU?

[Alexey Pasumansky]

Hello kirk,

This looks more adequate.

The performance may depend on the parameters used and image properties. Which quality option have you used and what is the image resolution?
Also I recommend to disable "use CPU" flag in the GPU Preferences tab, if it is enabled.

[kirk]

It's "ultra high" quality  and 575 images of Mavic pro drone each  3992x2242.   I need to build a super hi res mesh ( 100mil probably) to be then baked into displacement and normal map textures (UDIM unwrap) of a low res in-game terrain.     

 Not sure , perhaps I should do it in a few smaller chunks? But I have no idea how would I  deal with overlapping parts when will bake hi-res to low-res model in texture baking soft considering so huge poly count would be impossible to edit manually.

Wonder could I do it some other way without so huge mesh and dense cloud calculation?
 I see I can export camera based depth images and world space (maybe?) normal map for each camera. 

A crazy idea  but  should I  try to just use them as a replacement of color images for building low poly model  texture.    I need a depth map ray traced along low poly mesh normals  and I bet the image depth is ray traced along corresponding camera vectors so it wouldn't work probably.   But the normal maps may work I guess.    Are they camera space and  different for each camera or world space? 

After several hours it stopped to use GPU at all and the "left" counter is  showing more and more.

[Alexey Pasumansky]

Hello kirk,

Dense cloud filtering is CPU-only stage and may take long time in case of excessive image overlap. The processing time for this step can be reduced by setting up main/dense_cloud_max_neighbors Tweak via Advanced preferences to something about 50-60.

However, I assume that you may experience "not enough RAM" problem, when trying to build mesh from Ultra quality cloud for this project on 32 GB machine.

I could suggest to try the GPU-supported depth maps based mesh reconstruction option (thus skipping the dense cloud reconstruction option).

[kirk]

Thanks Alexey, 

Is that "dence_cloud_max_neighbor" setting  available only in Pro version? I have Standard one.  Or should I just add such a line into Tweak dialog?

With that Depth based method should I keep "re-use depth maps"  on or off?

[Mak11]

Kirk,

Just ad that line in the Tweak dialog, hit apply and then restart Metashape (note that this tweak only apply to the Dense Cloud generation process).

If you already have the depth-maps generated just re-use them for the new mesh method.

Mak

[kirk]

Thanks, Mak11