Slow Down on Mesh Creation

[nblachford]

Hey There,

Our organisation is interested in creating models of remote communities in the arctic. We have begun our pilot project by flying a drone around a number of our locations.
Intention is to have some large models of buildings (usually around 100) sitting on their natural topography. And then be able to take measurements and stats from our captures.

I still have lots to learn with the software but so far from my current tests we seem to be getting useful results. However I've a hard time finishing lots of the models. Some of my larger tests have worked out well, yet others have been freezing or appearing to take longer then I would have expected, mostly at the photo consistency based optimisation stage of mesh creation. While "preparing neighbours" zephyr "initializing". Smaller meshes have certainly passed through this stage quite quickly. Others such as this attached log file have done the first few hundred photos in a short period of time then slowed down to start giving me a photo every few hours.


The computers:
Dell Inspiron Laptop: [this computer is the one I leave on all night]
Windows 10
Zephyr Pro 4.002
CPU: I7 - 7700HQ 4 core
RAM: 16GB (1 module)
HD: 128GB SSD [ 30 GB free, produced the point clouds on the cache, however, the mesh stage, despite clearing, says that 30gb is too small, so the cache is on the HD]
2TB traditional HD
GPU: GeForce GTX 1050 Ti

Alternate computer [ more of my workspace ]
Microsoft Surface Studio:
Windows 10
CPU: I7 - 6820HQ 4 core
RAM: 32GB
HD: 2 TB traditional HD
GPU: GTX 980 M 4GB

Previously I have successful produced a dense cloud of a slightly larger community at 1.0 FPs using 2600 images. However, I was getting very stuck at the mesh creation stage.
Same issue, "preparing neighbours, initializing". The log shows lots of progress right away slowing down to an image every few hours. So I moved on to try again for this project.

This project was made from about 1000 images taken from about 30 mins of 4K video footage. Separated at 0.5 FPS.

This log file's Dense Cloud took about 1 week of the laptop running straight, using Aerial Default settings. It looked quite successful, the landscape looks good and the buildings all appear. I cleared off extra points that were unneeded and moved to the mesh stage.

During this stage "preparing neighbours" the Ram is always full, and the GPU and CPU are near idle.

So the question is how to I complete some of these mesh procedures? Will turning down a particular setting help to finish the operation? Are the machines just inadequate for the process?
At what point in the "preparing neighbours" stage should I give up? It's going right now through an image every hour or so with 500 still to go.

Thank you for your time,

Nick

[Roberto]

Hi Nick,

thanks for sharing the log. From the log, it seems that the RAM usage is more than the available on the system during the photoconsistency phase. As a consequence, Windows starts swapping from HD to RAM, and the processing is slowed down a lot.
You can switch to advanced options during the mesh creation and turn off the photoconsistency based optimization.

However, since I've noticed you are using v 4.002, could also please try to upgrade to the last version of Zephyr 4.009? I think we had addressed a similar issue and decreased the memory usage:

https://www.3dflow.net/3df-zephyr-pro-download-page/

Please let me know if it solve the issue or we can investigate further

[Andrea Alessi]

models of remote communities in the arctic.

this sounds amazing, when you complete the project i'd love to see your results (i love the arctic!)

[nblachford]

this sounds amazing, when you complete the project i'd love to see your results (i love the arctic!)

This may take awhile.
I'm certainly excited to show people. This is really a pilot project, so I don't expect to have the most professional results possible. We wanted to see what was possible working with available consumer level technology while piggy backing this project on another. Certain things don't work as we expect in our part of the world. For example, GPS signal for our DJI drones is on the scary side of unreliable [maybe because the magnetic south pole is south of me] so a lot of the flight paths really needed improvement.

I've taken the advice of updating Zephyr. to 4.009

I also took the advice of turning off photo consistency optimization. I aborted the process from the log file in my last post [data set of 1000 imgs] and restarted it without the photo consistency. It produced a mesh in about an hour. A couple screen shots attached. There is quite a bit of noise. I'd previously tried smaller samples of this community [ie: just the landfill] using the photo consistency and the meshes are a bit more articulate.

So I'd like to try the mesh generation again see if I can clean up a bit of the noise.

Following this tutorial https://www.3dflow.net/tutorial-poin...-mesh-filters/ : The settings "Hierarchical subdivisions 2, image resolution 50, iterations 40" appears to be where I should start?

If I keep having projects slow down like in the previous log file, when do you think its appropriate to abort and try with different settings?

[Andrea Alessi]

Hi Nick,

photoconsistency helps a lot (here is also a videotutorial by the way https://www.youtube.com/watch?v=k4z-dTHAg58 ) but is pretty intensive, so if your computer is struggling i would consider improving the hardware.

That said, also consider image quality and camera orientation. Noise may also be given for example if you have some cameras not oriented correctly. Did you use GCP ? Even if not, try placing a few control points and see if the epipolar line goes on another image where you expect it to be.

If you want to share images and zep we can happily have a look

[nblachford]

Hey There,

Thank you guys very much for the good advice.

I didn't use GCP at all. Would that have been an advantage?

Sharing the images/zep is a bit tough for me. The data sets in this project are around 20GB each (i have a couple balls rolling). Our northern internet is very slow, so I'll have to do that as a last resort if I am getting really frustrated.

Removing the photo consistency certainly shortened my processing time from 1 week approaching infinity to less than a day.

Using PBO as a mesh filter, after creation has also been successful after a couple tries. More RAM seems like it would help. I do get more articulate results after using the PBO filter, so in some situations I found I do want it.

Moving on I have a mesh I am satisfied with now. New problem now. The textures and mesh color data have become messed up. I had taking my first non-PBO mesh and textured it and had reasonable textures.

I filtered the mesh with PBO, BiLat Smoothing and Filled Holes. Now the color data is very wrong, even after filtering to update colors. Producing a texture map even to the original mesh is now having issues.

I've attached 2 images. You can see the colors of the houses have been painted with ice.

What did I do to destroy the texture data?

Thanks,

Nick

[Andrea Alessi]

Hi Nick,

GCP can be use to georeference your dataset - if you need - as well as set constraints to drive zephyr to the correct reconstruction, improving accuracy in some cases.

What do you mean by "have become messed up" ? do you mean in zephyr, or in export, or something else ? Unfortunately i don't see any image attached, can you please post the screenshot again?

Thank you!

[nblachford]

Oops sorry about that! Somehow I missed attaching them.

The meshes are the same between screenshots. It just appears to have started sampling the wrong cameras for the texture. It became inaccurate basically everywhere. This is all inside of Zephyr.

I have changed computers between building the first textured mesh and this second one. I'll try testing it today to see if that caused my issue.

Best,

Nick

[Andrea Alessi]

Hi,

i think that you simply picked the wrong photos. Zephyr looks for the photos in the same directory where the .zep is located, but if it can't find them, it will look up the absolute path where the images were saved.

So i think this happened. On your first computer your project probably looked like this

c:\users\myuser\documents\photos\project.zep
c:\users\myuser\documents\photos\flight1\img00.jpg
c:\users\myuser\documents\photos\flight1\img01.jpg
c:\users\myuser\documents\photos\flight2\img02.jpg

then you moved your project to another computer

c:\users\otheruser\documents\something\project.zep
c:\users\otheruser\documents\something\flight1\img 00.jpg
c:\users\otheruser\documents\something\flight1\img 01.jpg
c:\users\otheruser\documents\something\flight2\img 02.jpg

then the zep file could not find the images because they were neither in the root directory of the zep nor in the absolute path (which changed).

So zephyr pops up a window that tells you to select which photos are which - very easy to mistakenly set a wrong directory or even worse, change an image.

That's my guess, please double check and let me know!

[nblachford]

Yes, that appeared to be the issue after all. That was a pretty easy fix.

Thank you very much for the smart answer.