Airbus Ship Detection Challenge

Github Pages: https://puechtom.github.io/airbus_satellite/

Kaggle competition: Airbus is excited to challenge Kagglers to build a model that detects all ships in satellite images.

The dataset is composed of 208 162 images with a resolution of 768 by 768 pixels. The dataset is splitted into 192 556 training images and 15 606 testing images. Example of satellite images:

For training purpose, each training image has a corresponding segmentation mask. This mask will be used as ground truth by the network. The masks corresponding to the example images above:

U-Net architecture

I used a well known segmentation network, implemented with Keras in Python, in order to detect ships. The network architecture is described as below:

Results

Since I do not have access to a powerful GPU, I can not properly train the network. Nevertheless, it acheives a true positive detection rate of almost 74%.

From left to right, original images, ground truth mask and predicted mask by the network: