UNet node for image segmentation

[kluge7]

Motivation

Currently, testing a UNet segmentation model in ROS is either done through this launch file which relies on the full Isaac ROS TensorRT pipeline and requires a GPU-enabled setup with CUDA and TensorRT installed.
Alternatively, it’s done through this script, which hardcodes parameters, lacks a config/launch setup, and mixes inference and visualization logic (poorly written).

Description

We want to build a simple ROS2 node for running UNet segmentation models. The node should subscribe to a color image topic, run inference using a specified model, and publish both the segmentation mask and an overlay image for visualization. It should have a launch file and a config file.

Task

1. Implement the ROS2 node in vortex-deep-learning-pipelines

• Use ros2cli for making the package
• Subscribe to a color image topic
• Run UNet inference
• Publish segmentation mask and overlay image

2. Add config file

• Define parameters (model path, topics, resize size, threshold, device etc.)

3. Add launch file
4. Test functionality

• Verify on local setup
• Use Simulator or ROS bag for validation

5. Documentation

• Add README on how to launch and configure the node

Deliverables

• ROS2 node: Functional node that performs UNet inference, publishes both the segmentation mask and overlay image.
• Config file
• Launch file
• Test verification
• Documentation

References

• Pytorch-UNet - https://github.com/vortexntnu/Pytorch-UNet
• Basic UNet Ros node - https://github.com/vortexntnu/Pytorch-UNet/blob/master/predict_ros.py
• ros2cli for making package - https://github.com/vortexntnu/ros2cli
• Launch file example - https://github.com/vortexntnu/software-learning-period/blob/main/tutorial_packages/simple_publisher/launch/simple_publisher.launch.py
• Config file example - https://github.com/vortexntnu/software-learning-period/blob/main/tutorial_packages/simple_publisher/config/simple_publisher_config.yaml
• Similar node (use as example) - https://github.com/vortexntnu/vortex-yolo-inference

Contacts

@kluge7
@jorgenfj