The goal of the task is to create an algorithm for segmentation of ships in satellite images. Using this algorithm, for example, it is possible to detect contraband by comparing recognized ships with the online base of GPS coordinates of registered ships. As, for example, in medicine domain, the minimization of the False Negative is also more important in this task. It's easier to recheck every positive prediction for correctness instead of missing any of it. Accordingly, F-score over IoU thresholds with β=2 metric was selected for evaluation (weights recall higher).

Since predicting at least one pixel as a ship in an image without ships, we get a score of 0 for this picture, our algorithm must accurately understand if there is at least one ship in the picture. To do this, we will train a classifier that will predict the presence of at least one ship in the picture. This will reduce the False Negative errors and also save the time of the prediction, since we will not use segmentation network on pictures that are marked by the classifier as 'non-ship' images. Unet architecture was chosen for the segmentation network. Resnet34 for the classifier and Unet34 for the segmentation network were chosen. Accordingly, we can use our trained classifier network as pretrained encoder part for the Unet network instead of using Imagenet weights. Unet was first trained on only ship images (choosing only crops with ships) for faster training. Then dataset was balanced as approximitely 0.75 images w/ ships and 0.25 w/o ships for further training.

Submission:

This was achieved with training segmentation network for only ~400k images at total and having classifier with 97.5% accuracy only (256x256 resized prediction)

For training classifier Resnet34 network from https://github.com/qubvel/classification_models (need to installed for cls. training) repo was used and https://github.com/qubvel/segmentation_models fot the Unet34 segmentation network. For training the classifier you need to install 'classification_models' repo (described above). On the contrary, for training the segmentation model, the code for building the network is in the notebook.

Requirements.txt file was created manually and only for prediction notebook.

pip install -r requirements.txt

Download trained models Classification: https://drive.google.com/open?id=1yosonQpB8IQ7SfdX8_V7n8vOgYZ_JqtK . Trained for few epochs on 256 resized images, 97.5% accuracy achieved on default unbalanced dataset. Segmentation: https://drive.google.com/open?id=1PSgK4hySLrL7qILVqzwW1IKSDKPneuCf . Trained on 256x256 crops for few hours (1-2 epoch total because very low I/O ops. on the Kaggle Kernels) and finetuned on 384x384 crops, then for less than epoch on the fullsize 768x768 images.

Tehre are ShipSegmentor class in the 'Predict.ipynb' notebook that can be used both for prediction and visualization. For prediction create instance of this class with specified paths as in the example provided in the notebook. Then run code below for creating submission file. For this you also must download the trained models and specify paths for them. File has_ship_or_not already has classifier prediction so you can use it instead of classification from scratch.

ship_segmentor.predict_ship_masks_to_csv()

Visualize both train and test images prediction using ShipSegmentor class.

ship_segmentator.visualize_ship_segmentation(n_images=16, is_train=False, shuffle=True)

• Train fullsize input classifier and segmentation network for 30+ hours will give 0.72+ public score.
• With longer trained classifier we need train segmentation network with freezed encoder longer. Also with better classifier we can reduce 'no ship' images in dataset while training segmentation network for better performance and faster training.
• Also it's better to tune 'ship'/'no ship' images ratio using smaller crops and resnet18 as a simplified task for time saving.
• For better augmentation we can try train IRNet (https://github.com/jiwoon-ahn/irn) or other network for obtaining accurate ship segmentaiton masks (instead of rectangular). Then cut ships using this masks and past on 'no ship' images as a backgroung because we can assume we have 2d objects. This type of augmentation can make dataset much richer. Also then we can make more samples with close ships for better splitting of them.
• It was also found that in this task, the Adam optimizer often accumulates too much momentum, since after several epochs the error began to rise. This was solved by restarting the optimizer to reset the momentum which helps, but it might be better to configure the SGD optimizer.
• Pseudo-labeling