Airbus Ship Detection Challenge

This repository contains segmentation solution to the Airbus Ship Detection Challenge.

Requirements

To run the code in this repository, you need the main following dependencies (others are in requirements.txt): it is preferably to install python 3.8.10, the code is tested for this version of python. OS: ubuntu 20.04 LTS

tensorflow==2.8
opencv==4.8.0
numpy
pandas
matplotlib

You can install the dependencies by running the following command:

pip install -r requirements.txt

EDA

EDA notebook is located in notebooks directory. The color statictics, ship analysis was done in EDA.ipynb.

Weights and data

Download the dataset from the Airbus Ship Detection Challenge page on Kaggle.

Download the custom UNet model weights file from the given link and Unet from segmentation-models framework: link

Train

Download and extract the dataset files into the directory $PROJECT_PATH/data/.

Train the ship detection model by running the training script:

python src/train.py --output_weights=./output_directory/final_weights.h5 --checkpoint_dir ./checkpoints/ --image_size 768 768 --batch_size 16

Evaluate the model on the test set by running the test script (the weights can be both from custom model custom_unet_weights.py and library model unet_library_weights.hdf5):

python src/test.py --image_size 768 768 --batch_size 1 --weights ./weights/unet_library_weights.hdf5 --test_dir data/test_v2

To use the model directly as inference, run the following script (replace pathes as needed)

python src/inference.py --image_path ./data/train_v2/1c40bdeaa.jpg --weights ./weights/unet_library_weights.hdf5 --output_path ./predictions/1c40bdeaa.jpg

Examples

original image	mask	prediction