Computer Vision Image Segmentation Project

Overview

This project explores various image segmentation techniques on the Oxford-IIIT Pet Dataset. The goal is to implement and compare different segmentation approaches, ranging from traditional methods to deep learning models, while investigating how pre-training and feature extraction strategies impact segmentation performance.

---

Project Structure

• 📂 config: Configuration parameters
• 📂 data: Dataset loaders and augmentation utilities
• 📂 scripts: Scripts for preprocessing and running experiments
• 📂 utils: Utility functions for dataset processing
• 📂 Dataset: Original dataset

---

Generated Directories (not in version control)

• 📂 Processed_Dataset: Standardized dataset (generated via scripts)
• 📂 Augmented_Dataset: Dataset with augmentations (generated via scripts)

---

Environment Setup

This project requires several dependencies that can be installed using the provided environment file. Follow these steps to set up the environment:

conda env create -f environment.yml
conda activate cv-seg

---

Reproducibility

For consistent results, all data preprocessing includes a fixed random seed (RANDOM_SEED=42) to ensure:

• Train/validation splits are consistent
• Data augmentations are reproducible
• Models initialized with the same weights
• All generated datasets are created within the repository directory structure.

---

Task 1: Dataset Preprocessing and Augmentation

Goal

Prepare the Oxford-IIIT Pet Dataset for image segmentation by:

• Standardizing dimensions.
• Implementing data augmentation techniques to improve model training performance.

Dataset Overview

The original dataset contains images of cats and dogs with corresponding segmentation masks. The masks include:

• Class 0: Background pixels.
• Class 1: Cat pixels.
• Class 2: Dog pixels
• Class 3: Boundary/outline pixels.

---

Implementation

1. Dataset Standardization

We standardized the dataset using the following steps:

1. Resize Images: All images and masks are resized to 256×256 pixels.
2. Train/Val/Test Split: The dataset is split into:
   • Train (80%)
   • Validation (20%)
   • Test sets.
3. Mask Processing: Nearest-neighbor interpolation is used to preserve label information during resizing.

The implementation is available in preprocessing.py, which provides:

• standardize_dataset(): Handles image/mask resizing and splitting.
• get_train_val_split(): Creates training/validation splits using scikit-learn.

2. Data Augmentation

To improve model generalization, we implemented augmentation techniques using the Albumentations library:

• Geometric Transformations (applied to both images and masks):

  • Random flips (horizontal and vertical).
  • Random 90° rotations.
  • Elastic transformations.
  • Grid distortions.
  • Random cropping.

• Pixel-Level Transformations (applied to images only):

  • Brightness/contrast adjustments.
  • Gaussian noise addition.

For each training image, 3 augmented variants are generated, effectively quadrupling the training set size.

The augmentation functionality is provided in augmentation.py:

• get_training_augmentation(): Returns the augmentation pipeline for training data.
• get_validation_augmentation(): Provides minimal processing for validation data.

---

Usage

To preprocess the dataset, run the following script:

python scripts/prepare_dataset.py

This pipeline:

1. Creates a standardized dataset with consistent dimensions.
2. Generates augmented versions of all training images.
3. Displays dataset statistics.

---

Data Loading

For model training, we provide the PetSegmentationDataset class, which integrates seamlessly with PyTorch for efficient data loading.

---