OpenCV with cuDNN and DLAs engines

Hi, I’m new on ML and before start to learn, I need design and build a target learning platform, a robot. At the moment I just going to use examples from the nVidia ecosystem (as jetson-inference) but the robot AI subsystem must be at least planified now, as need make the camera setup etc.

I have a few answers I couldn’t find answer probably by don’t search properly, with a yes or not is enough:

  • Do DLA engines only can be used with TensorRT or can they be used by OpenCV through Cuda or a module as cuDNN?
  • Can OpenCV run TensorRT models direcly?
  • Can be used both DLA for the same model or they just can run a independent model.
  • Can I run a model with OpenCV and cuDNN (as Yolo or SSD) and other one or two models in DLA?
  • Can I run different models in a same video input?

Thank you!

Hi,

1. Only TensorRT.
2. The possible follow like this:
OpenCV CPU buffer -> memcpy to GPU -> TensorRT
OpenCV GPU bugger -> TensorRT
But you cannot run TensorRT directly with OpenCV interface.

3. No. You can only use DLA or DLA+GPU.
4. YES.
5. YES.

Based on your question, it’s recommended to check our Deepstream SDK first:

Thanks.

thank you so much Yeah, I will go deeper in TensorRT and Deepstream, this questions are to build a robot in base to the Xavier NX fatures capabilities. In example your answer saved me of buy a extra camera and helped me to can activate CUDA with my realsenses. thanks a lot, really.