Meeting date: 30-03-2023 (10h30)

Summary

After the last meeting Pedro asked me to get the Jetson ready to do inferences in DeepStream. The goal was to use models like YoloV8 and YoloV7 with trackers (DeepSORT and NvDCF). Turns out that to be able to do inferences with this models it is necessary to convert the model to a specific format and it is needed to run a specific script to generate the required config file. To run this scripts I had to install both PyTorch and TensorFlow and solve some problems with Pedro’s help. The tracker DeepSORT was a bit trick to get to work, but it is now very easy to use. Until now, I tested YoloV8s, YoloV8m and YoloV8l and I’m waiting for Pedro to finish what he is doing with YoloV7.

I worked bit more in the code that I started in the last week:

• I solved a problem that I was having when changing the input image dimensions

• I fixed some flaws, that would cause problems when using single task models

• I added an argument parser to chose the model and the module in a launch file

• I changed the format of the bounding boxes corners coordinates from float32 to int16

• In the receiver node, I fixed a problem with the labels and bounding boxes synchronization and now it is also plotting every thing that comes in the messages (Its not doing the synchronization between the inference messages and the original image: just plots as soon as it receives a new message)

  

From what I saw in the documentation (http://wiki.ros.org/Names and http://wiki.ros.org/roslaunch/XML/group) it is possible to add namespaces in the launch file and in the node file. I think that the better way is to do it in the node file, since it’s already needed to give to it both the model name and the source name as arguments. In the node file it’s only possible to add namespaces to topics

Objectives

• [x] Get the Jetson ready to perform inferences in DeepStream
• [x] Learn how to use new YOLO models in DeepStream
• [x] Learn how to use NvDCF and DeepSORT in DeepStream
• [x] Perform inferences with some models provided by Pedro Azevedo
• [x] Finish the scripts started in the last week
• [x] Find a way to create a Ros namespace based in the image source and model name (for a possible automatic subscription of nodes)
• [ ] Get the Ros node able to run in parallel with similar nodes