---

Learning Image Segmentation

This repository documents my journey into mastering image segmentation using deep learning. I've worked on several real-world challenges, implementing segmentation architectures like U-Net from scratch and applying them to competitive datasets.

📚 Projects

1. UNet From Scratch

Implemented U-Net architecture using PyTorch from the ground up.

• Built without high-level segmentation libraries.
• Includes training loop, data augmentation, and metric evaluation.
• Designed for educational purposes to understand every part of the pipeline.

2. Carvana Image Masking Challenge

Applied U-Net to a real-world car segmentation task from Kaggle.

• Dataset: Carvana image segmentation competition.
• Focus on binary mask generation (car vs background).
• Emphasis on preprocessing, model training, and postprocessing of masks.

3. Airbus Ship Detection Challenge

Tackled ship segmentation problem with complex, noisy satellite imagery.

• Dataset: Airbus Ship Detection from Kaggle.
• Includes techniques for handling missing masks and noisy labels.
• Improved postprocessing with connected component analysis.

🧠 Learning Goals

• Understand the U-Net architecture deeply.
• Learn how to train and evaluate segmentation models on different datasets.
• Explore techniques for data augmentation, loss functions (e.g. Dice, BCE), and postprocessing.
• Gain hands-on experience with real-world segmentation challenges.

🔧 Future Plans

• Add experiments with other architectures (e.g., DeepLabV3+, PSPNet).
• Try different loss combinations and training tricks.
• Explore multi-class segmentation datasets.