Meeting date: 02-03-2023 (10h30)

Summary

What was done

I searched about image segmentation in deepstream and about the integration of deepstream in ROS and faced some difficulties. Then I started to look for alternative ways to do inference tasks with high performance and found out a package, ‣, but this will be problematic to use with multitask neural networks and will take lots of time to adapt. I then decided to go for a simpler solution, but I had problems with Jetson that I have to solve first. Since I don't have time to solve all problems until the next meeting, I tested a multitask neural network (YOLOPv2) to have at least one thing working.

I would like to try to test all alternatives with a model compatible with all methods (deepstream, jetson-inference and pytorch-tensorrt) to see the difference in performance and conclude if is necessary to invest time in adaptations

Problems to solve

Jetson OS

Possible solution

Re-flash

Objectives

• [ ] Deepstream ROS
• [ ] Deep-learning Network for semantic segmentation

Daily report

24-02-2023

What was done

I searched for image segmentation in DeepStream and there is what I found out:

• It was not possible to do back in 2019: https://forums.developer.nvidia.com/t/image-segmentation/69991
• There is a NVIDIA repository: https://github.com/NVIDIA-AI-IOT/deepstream-segmentation-analytics, but it has not been updated for two years