Segmentation Model with Pre-trained MobileNet Encoder

Introduction

The objective is to design a custom decoder that predicts segmented masks for the given dataset. The dataset consists of 900 training images with corresponding masks, and 379 test images with masks.

Dataset Pre-processing and Custom Dataloader

• The images are resized to 128x128 during pre-processing.
• Custom dataloaders are implemented to load the dataset efficiently.

Experiments

Experiment 1: Feature Extraction

• Utilize a pre-trained MobileNet encoder trained on ImageNetV1.
• Design a custom decoder atop the encoder for the segmentation task.
• Train the decoder while keeping the encoder frozen.
• Options to take input for the decoder from the last layer of the encoder or from multiple layers are explored for potentially better results.

Experiment 2: Fine-tuning

• Keep the same architecture as in Experiment 1.
• Train the segmentation model while fine-tuning encoder weights.

Tasks

Perform the following tasks for both experiments:

1. Intersection over Union (IoU) and Dice Score: Calculate and report IoU and Dice score.
2. Loss Plots: Plot training and validation/testing loss curves.
3. Results Analysis: Analyze the results and provide observations.
4. Visualization: Visualize several samples alongside their generated masks and ground truth.
5. Comparative Analysis: Conduct a comparative analysis for both experiments.

How to Use

1. Clone this repository.
2. Download the ISIC 2016 dataset and place it in the data/ directory.
3. Run the main.ipynb files to train the segmentation models.