This is an implementation of the Kaggle competition Airbus Ship Detection Challenge | Kaggle .

The project include:

• training process with Unet or FPN, which is implemented by GitHub - qubvel/segmentation_models: Segmentation models with pretrained backbones. Keras.
• data augmentations implemented by GitHub - albu/albumentations: fast image augmentation library and easy to use wrapper around other libraries
• cycle learning rate (CLR) implemented by GitHub - bckenstler/CLR
• test procedure with TTA and the linear search of the best threshold

• 2019.2.22 first commit
• 2019.3.2 modify the data generate, change the demo_test filetype from .ipynb to .py, change the training procedure

• python==3.6.7
• Keras==2.1.2
• tensorflow==1.4.1

Run ./demo_train.py, then you will get the segmentation model in the “output_” + current time fold.

Run ./demo_test.py, finally you will get the submission file submission.csv in the “output_” + current time fold.