The goal of the challenge on Kaggle platform is pixel-wise semantic segmentation of salt bodies depicted on a seismic reflection images.

Here you can find a description of the 14th place solution by Argus team (Ruslan Baikulov, Nikolay Falaleev).

The final result of participation: the 14th place out of 3234 teams (top-0.5%, Kaggle gold megal).

Contents:

1. About Kaggle TGS Salt Identification Challenge
2. Semantic Segmentation of Seismic Reflection Images

Sample results

Example of the whole mosaic post-processing. Green/blue - salt/empty regions from the train dataset; red - predicted mask; yellow - inpainted by the post-processing (used in the final submission).

The project code is available on Github.

The title image is from here.

Competition posts