nlztrk / kaggle_identifying_structural_damages

Identifying Structural Damages with Image Segmentation

This notebook was created for use in Anadolu Sigorta Datathon. The competition awarded the solution with fourth place on the basis of the score, and the third place when evaluated together with the presentations.

Solution Pipeline

• Image Augmentation
• Customizing weights for class-imbalance
• Pre-trained backbone (DeepLabV3+)
• Pre-training on minority classes
• Full-training with 3 seperate model on full-data
• Blended prediction

Running the Whole Process

You can run the notebook file to execute full process.

Please cite this repository when using the content:

@misc{ozturk2020_anadoludatathon,
author = {Ozturk, Anil and Gunel, Mustafa Burak},
title = {Identifying Structural Damages with Image Segmentation},
year = {2020},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/nlztrk/kaggle_identifying_structural_damages}},
}