Xplendia - Off-Road Semantic Segmentation Dashboard

This project is a Streamlit-based hackathon demo for off-road semantic segmentation. It combines a polished dashboard with a PyTorch segmentation backend so users can upload an image, run inference, view the predicted mask, and compute IoU when a ground-truth mask is available.

What It Does

• Upload an image and run segmentation
• Show the prediction mask in clean class colors
• Show an overlay view for easy visual comparison
• Compute IoU when a ground-truth mask is provided
• Display model performance, risk analysis, and summary pages
• Fall back safely if the trained model is missing

Project Structure

• project/app.py - Streamlit dashboard
• project/iou_docs.py - IoU explanation block shown in the dashboard
• model_loader.py - Model loading, preprocessing, prediction, mask coloring, and IoU helpers
• XPLENDIA_GitHub/segmentation_head.pth - Model checkpoint used if a local one is not found

Requirements

Core packages:

• streamlit
• numpy
• pillow
• torch
• opencv-python

Optional packages:

• pandas
• matplotlib

The app works even if the optional packages are missing, but some charts will fall back to simpler display modes.

How to Run

1. Open a terminal in the project root.
2. Activate your virtual environment if you are using one.
3. Install dependencies:

pip install -r requirements.txt

4. Start the app:

streamlit run project/app.py

IoU

IoU stands for Intersection over Union.

Formula:

IoU = Intersection / Union

The dashboard shows a dedicated IoU explanation section and uses the same calculation in the backend when a ground-truth mask is available.

Expected Input Files

For the sample viewer and risk analysis pages:

• input image files such as sample1_input.png
• ground-truth mask files such as sample1_mask.png
• prediction files such as sample1_pred.png

For live upload:

• upload an image
• optionally upload a matching ground-truth mask

Notes

• The dashboard keeps its dark theme and layout.
• The backend tries to load the real model first.
• If the model is missing or invalid, a safe fallback model keeps the app usable.
• IoU is only shown when a ground-truth mask is available.

Troubleshooting

• If the app opens but IoU says Not Available, upload a ground-truth mask too.
• If the model is missing, check XPLENDIA_GitHub/segmentation_head.pth.
• If a package import fails, reinstall the dependencies in a clean virtual environment.

Submission Checklist

• Dashboard starts successfully
• Upload and prediction flow works
• Segmentation mask is shown
• Overlay view is shown
• IoU section is visible and documented
• Model fallback works if the checkpoint is missing