The Nvidia L4T APT repo provides the libopencv and libopencv-python packages, which are compiled for the Jetson.
But this version of OpenCV is built with support for neither OpenCL nor CUDA, which undermines the key reason for using Jetson.
I understand that this can be compiled and installed manually, but can Nvidia please consider publishing an accelerated OpenCV as this is certainly a commonly used library according to these forums?
Refer this thread;
For Jetson TX2,
-D CUDA_ARCH_BIN="6.2"
Thanks for that pointer (and the correct CUDA arch to provide, which I didn’t know), but I hope Nvidia considers publishing a version of OpenCV that is built with the CUDA features that they themselves contributed.
That’s right.
Sometimes, I hope ^^
On the other hand, JetPack should support many platforms from TX1 to AGX Xavier.
But they have slightly different configuration each other.
I think that’s not easy to cover all.
Ive created some .dsc source package bundled for building 4.2.0, 4.3.0 and 4.5.1 with cuda enabled that built and installed on JP 4.4, but breaks nvidia-vpi because it uses OpenCV 4.1.1. (I also had to patch 4.2 and 4.3 to handle CUDNN v8). They’re based on existing .dsc packages with debian/control|rule modifications and I imagine I dorked some stuff. I’d be willing to share if people could help me make sure all the dependencies and debian packaging stuff works.
It would be great if you wanted to share your work. Can you create a GitHub (or similar) project to share these? If there are fixes I come up with I am happy to contribute back.
I’ll see if I can fork off the original repo and post up the ones I modified. Right now they’re just hacks of .dsc bundles.