Airbus Semantic Segmentation using CNN

Summary

This repository contains a project on semantic segmentation using airbus ship detection dataset.

• Task: Semantic Segmentation
  
• Data: Link
  
• Neural Network: U-Net like architecture
• Scoring function: Dice score (f1 score)

How to run:

• Download data
  • Download Kaggle Competition data
  • Download mask images (to save time) Link
• Prepare folder structure
  • Create data folder and extract Kaggle data into it
  • Create masks_v2 folder inside data folder. Extract mask images into masks_v2 folder
• Run eda.ipynb to create training dataframe

• To train

python train.py --backbone=mobilenetv2

--backbone - choose UNet encoder backbone
python train.py -h for more details

• To predict

python inference.py --download --visualize_inference

--download - Download pre-trained model
--compare_to_gt - Compare prediction to Ground Truth
--visualize_inference - Visualize 10 inference results
--predict_all - Takes a long time! Predict on all test images
--show_submission - Visualize random predictions from RLE encoded submission.scv
python inference.py -h for more details

Model results

Model info

Trained 30 epochs, on 5k images (4k training and 1k validation), batch size 10, using basic image augmentation to aid in training. Metric - f1-score, loss function - binary crossentropy.

Model training history

Model evaluation

Prediction vs ground truth

Inference example

Project Structure:

├───data # Data folder
│   └─── masks_v2 # RLE decoded masks
├───EDA
│   └───eda.ipynb # EDA and initial data prep
├───train.py # Define model
├───inference.py # Model inference 
├───results # Store model execution results
│   ├───inference # Model inference results
│   ├───validation # Model validation data predictions
│   └───submission # Images from predicted RLE
├───checkpoints # Folder with best model checkpoints
├───constants.py # Declare variables
├───data_prep.py # Create data generators and augment pictures
├───helper_funcs.py
├───metrics.py
├───.gitignore
└───requirements.txt

Kaggle Notebook

Used it to run all model training View

Conclusion:

Project was very challenging as I had 0 previous experience with neural networks and keras, but it was very fun nonetheless.