Image Segmentation Challenge

Welcome to the image segmentation challenge! This project provides a baseline pipeline using scribble-based supervision for binary image segmentation. You will work with a dataset of natural images, sparse scribbles indicating foreground/background, and optionally ground truth labels for evaluation.

---

Project Structure

.
├── challenge.py           # Main script for training and testing pipeline
├── util.py                # Utility functions (dataset I/O, KNN model, visualization)
├── dataset/
│   ├── training/
│   │   ├── images/
│   │   ├── scribbles/
│   │   ├── ground_truth/
│   │   └── predictions/   # Created by the script
│   └── test/
│       ├── images/
│       ├── scribbles/
│       └── predictions/   # Created by the script

---

Requirements

Install required Python packages via pip:

pip install numpy pillow matplotlib scikit-learn

---

Running the Baseline

Execute the following script:

python challenge.py

This script will:

1. Load the training dataset.
2. Segment images using a K-Nearest Neighbors (KNN) classifier (k=3) based on scribble labels.
3. Store the predicted segmentation masks.
4. Visualize one randomly selected result.
5. Repeat the process on the test dataset (without ground truth).

---

Contents

challenge.py

• Loads training and test datasets.
• Performs segmentation using a baseline KNN model.
• Stores predicted masks to predictions/.
• Visualizes one result for the training set.

util.py

A utility module containing:

• Dataset Handling
  • load_dataset(): Load images, scribbles, ground truths.
  • store_predictions(): Save predicted segmentation masks with color palettes.
• Model
  • segment_with_knn(): A KNN-based baseline using scribble supervision.
• Visualization
  • visualize(): Display image + scribbles, ground truth, and predicted mask.
• Evaluation
  • evaluate_binary_miou(): Computes mean Intersection over Union (IoU).

---

Scribble Supervision Format

• Scribble masks use:
  • 0: Background
  • 1: Foreground
  • 255: Unlabeled

Predictions and ground truths are binary masks using 0 and 1.

---

Tasks for Students

• Improve the segmentation performance beyond the KNN baseline.
• Try different models or distance metrics.
• Use the provided visualization and evaluation tools to test your ideas.
• Optionally, extend to multi-class segmentation (requires additional modification).

---

Questions?

Please contact the course staff or teaching assistants for help with setup, debugging, or project expectations.

Happy coding!