Computer Vision Python packages with Cuda-GPU support for Jetson AGX Xavier?

Dear Nvidia Collegues,

Which computer vision oriented python packages are available with Cuda-GPU acceleration directly for Jetson AGX Xavier devices? I know about Tensorflow, Pythorch and OpenCV. I would like to know the others.

I am presently working in a Computer Vision Research Project where we would use a Jetson AGX Xavier device for compute vision in python. We have already installed tensorflow on a Jetson AGX Xavier and it works fine in gpu mode . We would like to install and test the other available Cuda-GPU-Python packages for Computer Vision.

Hi @Tamas23, the other popular ones I am familiar with are cupy (numpy for CUDA), pyCUDA (run CUDA kernels from Python), numba (autovectorization for GPU).

I would check out the latest l4t-ml container for JetPack 4.6. (L4T R32.6.1). They have all of the above already installed. The R32.6.1 version of the container has OpenCV built with CUDA enabled.

Also, there is TensorRT for optimized DNN inferencing with it’s Python API. jetson-inference is a wrapper around TensorRT and various models and has a Python API. DeepStream has a Python API and uses TensorRT and supports multiple simultaneous streams/models

Hi,

Thank you very much for the detailed reply!
What about Dlib (dlib · PyPI) and Mediapipe(mediapipe · PyPI)?
I use this maching learning packages in CPU mode regularly in our lab.

Can you see any chance that Dlib and Mediapipe exists in GPU mode directly for Jetson AGX Xavier devices?

Although I haven’t personally used them before, I’ve definitely seen topics on the forums about people using DLIB with GPU acceleration on Jetson. Mediapipe as well.

Thank you, if I find a description for the installation of Mediapipe/Dlib for Jetson Nano/ Jetson Xavier NX/Jetson TX2
etc will this type of installation procedure also appropriate for Jetson AGX Xavier or every Jetson model has its own installation procedure for every GPU specific python package?

Hi,

Please install dlib from the source.
The same installation command can be used across different Jetson platforms.

Ex.

$ wget http://dlib.net/files/dlib-[ver].tar.bz2
$ tar jxvf dlib-[ver].tar.bz2 dlib-[ver]
$ cd dlib-[ver]/
$ mkdir build && cd build
$ cmake ..
$ cmake --build .
$ cd ../
$ sudo python3 setup.py install

Thank you we will try it!