I purchased the WebODM installer so I wouldn’t waste time or have to mess with system environment settings. I installed it in the Windows environment which would run natively.
The problem is that after installation I was unable to process the files because the process simply did not finish.
The computer I use normally processes the same photos in Pix4dFields.
Windows 11 Pro
16Gb RAM
Intel I7 10750H
GeForce GTX 1650
SSD 1Tb (500+ Gb free)
I believe that WebODM is not using the graphics card’s resources for image processing.
I hope someone can guide me through the necessary procedures to activate this processing.
From what I understand WebODM uses CUDA on the graphics card for some of its processing tasks and looking at this List CUDA GPUs - Compute Capability | NVIDIA Developer, the GTX 1650 isn’t a CUDA enabled GPU
The GTX 1650 is CUDA capable, but WebODM will fall back to CPU if the card can’t hold the textures in VRAM.
16GB will not be much memory, so you might have to resize your images to 12MP at the most during Task creation, especially if you are using more than about 250 images.
Please keep in mind that Fields is not doing a full 3D reconstruction and dense pointcloud like we do by default, so our pipeline is more comparable to Pix4d Mapper.
You can try using the Fields preset during Task creation to work a bit more like Pix4d Fields.
The texts are in Portuguese but in the lines below we have 4Gb of RAM dedicated for video and 12Gb total graphics memory
Relatório de informações do Sistema NVIDIA criado em: 10/30/2023 15:10:17
Nome do sistema: ACP
[Monitor]
Sistema operacional: Windows 10 Pro, 64-bit
Versão do DirectX: 12.0
Processador da GPU: GeForce GTX 1650
Versão do driver: 462.30
Tipo de driver: DCH
Nível de recursos do Direct3D: 12_1
Cores CUDA: 896
Clock do núcleo: 1515 MHz
Taxa de dados da memória: 12.00 Gbps
Interface da memória: 128bits
Largura de banda da memória: 192.03 GB/s Memória total de gráficos disponível: 12221 MB Memória de vídeo dedicada: 4096 MB GDDR6
Memória de vídeo dedicada: 0 MB
Memória compartilhada do sistema: 8125 MB
Versão BIOS do vídeo: 90.17.4B.00.1C
IRQ: Not used
Barramento: PCI Express x16 Gen3
ID do dispositivo: 10DE 1F99 3FA217AA
Número da peça: 4904 0011
[Componentes]
nvui.dll 8.17.14.6230 NVIDIA User Experience Driver Component
nvxdplcy.dll 8.17.14.6230 NVIDIA User Experience Driver Component
nvxdbat.dll 8.17.14.6230 NVIDIA User Experience Driver Component
nvxdapix.dll 8.17.14.6230 NVIDIA User Experience Driver Component
NVCPL.DLL 8.17.14.6230 NVIDIA User Experience Driver Component
nvCplUIR.dll 8.1.940.0 NVIDIA Control Panel
nvCplUI.exe 8.1.940.0 NVIDIA Control Panel
nvViTvSR.dll 27.21.14.6230 NVIDIA Video Server
nvViTvS.dll 27.21.14.6230 NVIDIA Video Server
nvDispSR.dll 27.21.14.6230 NVIDIA Display Server
nvDispS.dll 27.21.14.6230 NVIDIA Display Server
nvLicensingS.dll 6.14.14.6230 NVIDIA Licensing Server
nvWSSR.dll 27.21.14.6230 NVIDIA Workstation Server
nvWSS.dll 27.21.14.6230 NVIDIA Workstation Server
nvDevToolSR.dll 27.21.14.6230 NVIDIA Licensing Server
nvDevToolS.dll 27.21.14.6230 NVIDIA 3D Settings Server
NVCUDA64.DLL 27.21.14.6230 NVIDIA CUDA 11.2.162 driver
nvGameSR.dll 27.21.14.6230 NVIDIA 3D Settings Server
nvGameS.dll 27.21.14.6230 NVIDIA 3D Settings Server
4GB on card, and it can request to use some of your system RAM for up to 12GB.
CUDA texture dimension limitations are hard to track down, but resizing your images down will likely help.
Be careful with a subset of images that you select a contiguous chunk of images with good overlap, otherwise reconstruction will fail, regardless of how much resources the system may have.
I’m not sure that list is complete, my GTX 1650 Super isn’t on that list either, and it is certainly used for dense reconstruction, and feature extraction for resized images.
Regarding the GPU, I now captured it during the process and I realize that it is used, I think it is underused because when I use Pix4dFields it hits 100% for processing and I can process more than 150 photos.
I’m processing the images at 50% resolution, they were 2048 and I lowered it to 1024.
By reducing the image, it was possible to complete the work with 206 photos of the area to be processed.
Now how can I know which is ideal to maintain quality results?
Unfortunately the machine is already at its maximum capacity but I don’t see the memory or cpu or disk consumption or anything else getting close to its limit (100%). Only brief periods of the CPU reach this level.
The GPU being used is only from the Intel graphics card and not from NVidia.
I saw in another article that NodeODM does not use multiple GPUs. The alternative for now would be to perhaps disable the Intel one to use the NVidia one.
