# Downloading the data
First, you should accept the competition conditions. Then, after installing the Kaggle CLI, create a data folder
under asd and run the following command inside it:
kaggle competitions download -c airbus-ship-detection -o -w
unzip airbus_ship_detection/asd/data/test_v2.zip -d airbus_ship_detection/asd/data/train_v2
unzip airbus_ship_detection/asd/data/test_v2.zip -d airbus_ship_detection/asd/data/test_v2
# Baseline
Notice that, since the "no ship" class is predominate, submitting a file with empty values for the EncodedPixel
columns leads to a 0.847 score on the public leaderboard. This will probably be different in the private leaderboard.
It is thus necessary to account for this imbalance in order to get good performances.
# Submission
Generate the submission file using the run.py script. Then, use the Kaggle API tool as follows:
kaggle competitions submit -c airbus-ship-detection -f /path/to/submission/file -m "Model description message"
# Sanity check
ls train_v2 -1 | wc -l
should return: 192556
and
ls test_v2 -1 | wc -l
should return: 15606
# Data leakage
It appears that the test images are transformed images for the train ones (translated and cropped images). Thus, the test masks have been release. New test data should be available in the upcoming days (or weeks). For more details, check this discussion.
The new data is available and ends with _v2.
# Resources
- https://www.kaggle.com/c/data-science-bowl-2018/discussion/54741 => a great discussion post about advanced instance segmentation tricks. This is the winning solution for the data bowl 2018 challenge.
- https://www.kaggle.com/c/tgs-salt-identification-challenge/discussion/69291 => another great discussion post about the winning model for TGS competition.
- https://github.com/qubvel/segmentation_models => a great segmentation framework (pretrained weights and common architectures).
- https://github.com/ZFTurbo/ZF_UNET_224_Pretrained_Model => U-net architecture with pre-trained weights. To try later.
- https://arxiv.org/pdf/1801.05746.pdf => TernausNet model: U-net architecture with pre-trained VGG11 encoder. The github repo is here: https://github.com/ternaus/TernausNet.