UNET model for Airbus Ship Detection Challenge

Solution

Dataset for training and model inputs

• Large amount of dataset consists of images with no ships. So, were used 9000 images with ships, and 1000 without. We splitted 80 % for training, 20 % - for validation
• Due to large input images - 768x768x3, we model are using model with 256x256x3 input shape. This decision came from large model size, which we got from original inputs.
• To train model, we cutted input train image as 3x3 grid, and selected tile with the largest ship area in it's mask

Training

• Used BCE+DiceLoss
• Used Adam optimizer:
  • 8 epochs with 0.001 lr
  • 8 epochs with 0.0005 lr
  • 8 epochs with 0.0001 lr
  • 8 epochs with 0.00001 lr

Results

We got the best validation dice score - 0.8627

You can checkout Kaggle notebook with solution, where model was trained (Thanks, Kaggle!)

https://www.kaggle.com/code/volodymyrhryniuk/unet-for-airbus-ship-segmentation-challange/notebook