What I lose if I swap to OpenCV 3.4.x with Cuda and cudnn?

Will work Deelstream5, TensorRT, VPI 1.0, Isaac,Jetson-inference…?

As Jetsons are not updating to 20.04, but they are not keeping the native 18.08 and ROS opencV3 versions, in robotics with ROS practically everything get broken from one side (OS dependencies), or the other (opencv, cuda…), so it is almost impossible develop anything, in a productive way.

So what Nvidia software stuff will stop to work?

Hi @fpsychosis, I recommending trying this script to recompile OpenCV with CUDA/cuDNN enabled:

I haven’t tried downgrading, so I can’t speak to that directly, however jetson-inference shouldn’t be effected because it doesn’t require OpenCV. Regarding ROS/ROS2, you can still use newer version in container - you can find those through this page:

https://nvidia-ai-iot.github.io/ros2_jetson/

Or if you prefer not to use container, ROS Noetic/Foxy can be built/installed from source (which is what the containers do)

1 Like