Jetpack 4.6 unable to compile OpenCV 4.1.1,4.2.0, 4.4.0 with CUDA and cuDNN enabled and other bits

First, why JP4.6 use CUDA 10.2 and cuDNN 8 if comes with OpenCV 4.1.1 which cannot even be compiled against such versions and take their advantage? You killing the performance of my project and SBC with no reason, except probably in some way benefict you own alternative software as VPI or whatever.
To compile, in theory, you need install OpenCV 4.4, but the fact is that if you update at least via alt or you are in melodic, it will not compile. Tried 4.1.1, 4.2.0,4.4.0 and same kind of error.

I’m using the @mdegans script and it fails in any OpenCV version, (trying now 4.5).

I know a case that could compile with OpenCV4.4 but he is in Focal 20.04.

I was able to compile 3.4.14 with cuDNN and CUDA if that does mean something for anyone.

Why in every new jetpack there is some movement against your users?

Tired of in every new Jetpack my £500 SBC worth less and make me lose weeks or even months of work.

Obscure company

Please file an issue on GitHub.com for my script.

I will look into it. I haven’t had time to test the script on JP 4.6.

To help others, please post a link to the issue here as well.

Edited to add: I don’t think Nvidia does anything to sabotage OpenCV. It’s just not their product and not their responsibility to fix when the build system breaks somehow.

You are doing something wrong. I have compiled OpenCV 4.4 with Michaels script on my Xavier-AGX running JetPack 4.6.
I have made two changes in the configuration

  • -D CUDNN_VERSION='8.2' (instead of ‘8.0’, don’t know if that is necessary, though)
  • -D BUILD_opencv_python2=OFF (to turn of Python2 support as I don’t need it…

Nice ranting… but not very helpful…

Maybe your installation setup is different, i.e. I need python2, or you flashed a new image.

My adjetive is totally justified, i.e. the partneship camera program or the features advertised in release that never will be , or just stop to update jetson Nanos, for me that , other things, as realsense support in ISSAC is proper of a obscure company.

Thanks by confirm OpenCV 4.4 can run with CUDA and cuDNN

You script.works properly, I can install 3.4.14 or 4.5.x.
Thanks by your GitHub, share and try to help.

I agree is not responsibility of nVidia OpenCV maintenance , but it is their responsibility choose the right and compatible versions of the software that build they include in their software ecosystem and this take advantage of their Hardware, even if is of 3rd parties, as it is included by default.
Have CUDA GPU, with OpenCV that support it, and do not include compatible versions (4.1.1 is not), just have no sense.

Thanks a lot, really.

That’s a fair point. They should probably remove OpenCV entirely since it’s confusing and leads to missed expectations. Things might perform poorly and people could blame Nvidia.

Should be built by the script. Both versions are supported since the Nano ships with python2 as a default python.

IIRC, Realsense itself is EOLed. I doubt either Intel or Nvidia will bother much supporting a product that won’t be produced anymore. It’s a shame since it was a nice camera. There is a kickstarter doing something similiar but I can’t vouch for it.

That’s interesting. Does removing the line entirely work? IIRC I added that on the advice of @Honey_Patouceul because the version was not automatically detected. That may be fixed now. If it is, it can probably just be removed.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.