Airbus Ship Detection

Project derived from the Kaggle challenge "Airbus Ship Detection" to detect all ships in satellite images and locate them in every single image.

If you have more questions, feel free to contact [email protected] or emails listed here (Section A1, 09: Airbus Ship Detection).

Resources

Our project mainly consists of Model repo and Web repo. Issues and pull requests for these repos are welcome.

• In the Model repo, you will see how we build and train our model, thus detecting ships in testing set. (https://github.com/BUGenerator/Model)
• In the Web repo, you will see how we build our webpage and web server. (https://github.com/BUGenerator/Web)

Other resources are listed below (some need BU Kerbose account):

• Agile board: https://github.com/orgs/BUGenerator/projects
• Sprint 1 Slides: https://drive.google.com/open?id=1nW0XfrfmcUOdIXQVQkb_QdVCRu0mEHe2
• Sprint 2 Slides: https://drive.google.com/open?id=1hg6p04cKZ-N-WCua3hLS_kNUqgNbzFMCdizPbBcJbNw

Datasets

• Training set includes 231,724 satellite images.
  • For images in the training set, the ground truth EncodedPixels are provided in the train_ship_segmentations.csv file.
  • EncodedPixels is blank if there are no ships in the image.
• Testing set includes 15,606 satellite images
  • For images in the testing set, our goal is to predict the EncodedPixels.

Model

For this task, we compared several models and chose U-Net model for now.

U-Net is highly suitable for binary segmentation of images, which is exactly our task: binary segmentation for is or is not a ship.

Webpage

• For frontend, we build it with Bootstrap. For backend, we build it with Flask in Python.
• The web server is deployed on AWS Elastic Beanstalk. Thanks to Semaphore CI, we are able to auto deploy the program onto Beanstalk when new commits are pushed onto https://github.com/BUGenerator/Web/tree/master.

Usage

1. Check out the webpage

Visit our ship detection webpage at https://www.shipdetection.ml/.

Notice that the webpage is still a prototype. For now, the function including uploading images and outputing images can be done. Futher development are to be built.

2. Run locally

To directly train and run the ship detector locally, you will need to setup your environment. Please refer to https://github.com/BUGenerator/Model/tree/master/tools/ENG_Grid for a development environment. We will have instructions on simpler detection-only environment setup later.

After that, you can go through the .ipynb file by:

$ jupyter notebook

We will upload the model we have trained and the standalone detection module later. Sorry about this.