Isaac ROS is very useful tool for robotics engineers, but it looks like not all parts of the puzzle are presented.
I saw the webinar about yolov5 and it looks great, but where can I find any instructions (aside example) how to write the cpp decoder using NITROS?
I know that I can take the example from detectnetv2, but is there some place to find some instructions?
Thank you in advance
Glad you’re finding Isaac ROS useful! For writing your own cpp decoder, you have two options at the moment:
- Write one against ROS msg TensorList like any other ROS 2 node. This will incur an early copy of GPU memory back to CPU if you were going to process the inference output.
- Copy the GXF-based code for DetectNetV2 or UNet and mimic the setup to take advantage of NITROS. We haven’t provided much instruction there besides the code itself because of #3.
- Wait for an upcoming release where you can write your own C++/CUDA code and publish/subscribe through “managed” NITROS.
For now, I would go with option 1 as it is easiest to retrofit to option 3 if you’re not able to follow option 2. Otherwise, you can always ask specific questions on option 2 and we’ll do our best to help.
Can you help me to decide?
- When does the new release will be avaible?
- Will it help me for old versions like 5.0.2 (these are version my client has already in production)
Because if it comes soon and is back compatible, I’ll do as you say, take the option 1 and then rewrite it using 3. Otherwise, I have no choice to go for 2…