Fix Issue #245: Add UNet tutorial for DetectionMetrics

[vaibhavi089]

Add UNet Tutorial for RELLIS-3D Semantic Segmentation

Overview

This PR introduces a step-by-step UNet tutorial for the RELLIS-3D dataset, demonstrating how to train a semantic segmentation model using PyTorch. It is designed to help newcomers quickly get started with off-road scene understanding and model training.

Changes Included

• Added tutorials/UnetOnRellis3D.ipynb containing:
  • Dataset loading and preprocessing for RELLIS-3D
  • UNet model definition
  • Training loop with metrics computation (mIoU, Pixel Accuracy)
  • Example inference and visualization of results
• Fully commented code cells for clarity and educational purposes
• Ensures reproducibility with minimal setup

Why This is Useful

• Provides hands-on guidance for using RELLIS-3D for semantic segmentation.
• Helps students, researchers, and developers quickly implement a UNet baseline.
• Enhances the repo’s value by adding educational content, not just raw datasets.

Testing & Validation

• Tested on a sample subset of RELLIS-3D.
• Metrics computed include mIoU and pixel accuracy to ensure correct training.

Next Steps (Optional)

• Extend to other segmentation models (DeepLabV3, FCN) in future tutorials.
• Add visualization of loss curves and predicted masks for more interactive learning.

Request for Review

Please review this tutorial for clarity, correctness, and usefulness. Feedback is welcome to make it even more beginner-friendly.

[dpascualhe]

Hello there 👋

Thanks a lot for your contribution! For the tutorial to be merged there are some critical issues that should be addressed:

1. DetectionMetrics library is not used at all. It should be used at least to load the RELLIS-3D dataset using our Rellis3DImageSegmentationDataset, and more importantly, to evaluate the model which is the main functionality our library provides.
2. DetectionMetrics focus is model evaluation, not training. While it might be interesting to add a tutorial that do both, the focus of the tutorial should be the evaluation with DetectionMetrics.
3. Tutorials go in the examples/ directory.

Some resources for you to get started with DetectionMetrics:

• Installation instructions: https://jderobot.github.io/DetectionMetrics/v2/installation/
• Notebook tutorial: https://github.com/JdeRobot/DetectionMetrics/blob/master/examples/tutorial_image_segmentation.ipynb

[vaibhavi089]

Thanks for the feedback! @dpascualhe
I have moved the tutorial to the examples/ directory and updated it to:

Use Rellis3DImageSegmentationDataset for loading the dataset

Keep the UNet training part

Add evaluation with SegmentationEvaluator from DetectionMetrics

Please let me know if any further changes are needed.

[dpascualhe]

@vaibhavi089 your notebook does not follow the installation instructions for DetectionMetrics, nor does the usage align with the actual library (for example, SegmentationEvaluator is not a class that exists in DetectionMetrics). While we welcome AI-assisted PRs, it’s important that contributions are properly tested and working before submission. From the current state, it seems this notebook has not been tested.

If you’d like to contribute, I encourage you once again to go through:

• Installation instructions: https://jderobot.github.io/DetectionMetrics/v2/installation/
• Notebook tutorial: https://github.com/JdeRobot/DetectionMetrics/blob/master/examples/tutorial_image_segmentation.ipynb

to get familiar with the tool. Once you’ve done that, please revisit your notebook, ensure it uses DetectionMetrics correctly, and provide proof of it working locally. Otherwise, we’ll have to close the PR.

We look forward to your improved contribution!