Skip to content

ybabakhin/kaggle_salt_bes_phalanx

Repository files navigation

1st Place Solution for the Kaggle TGS Salt Identification Challenge (b.e.s. & phalanx)

Paper describing the solution:

Semi-Supervised Segmentation of Salt Bodies in Seismic Images using an Ensemble of Convolutional Neural Networks
German Conference on Pattern Recognition (GCPR), 2019
Yauhen Babakhin, Artsiom Sanakoyeu, Hirotoshi Kitamura
https://arxiv.org/abs/1904.04445

Kaggle post about the solution: link.

ENVIRONMENT

The solution is available as a Docker container. The following dependecies should be installed:

DATA SETUP

Download and unzip competition data into data/ directory. One could specify local path to the new test images in SETTINGS.json file (NEW_TEST_IMAGES_DATA field). The competition test data is used by default.

WEIGHTS SETUP

To get the weights from the final stage models download them from google drive and unzip into corresponding bes/weights/ and phalanx/weights directories.

DOCKER SETUP

To build and start a docker container run:

cd docker 
./build.sh
./run.sh

MODEL BUILD

  1. train models from scratch

    a) trains all models from scratch

    b) expect this to run for about 16 days on a single GTX1080Ti

  2. make prediction

    a) uses weights from the final stage models to make predictions

    b) expect this to run for 3.5 hours for 18,000 test images on a single GTX1080Ti

Commands to run each build are presented below:

1. train models (creates model weights in bes/weights and phalanx/weights)

./train.sh

2. make prediction (creates predictions/test_prediction.csv)

./predict.sh

ADDITIONAL NOTES

  1. Model weights are saved in bes/weights and phalanx/weights for b.e.s. and phalanx models respectively

  2. Individual model predictions before ensembling are stored in bes/predictions (lots of .png images) and phalanx/predictions (.npy files)

  3. Scripts to generate initial folds and jigsaw mosaics are located in bes/datasets: generate_folds.py and Generate_Mosaic.R

CITATION

If you find this code useful, please cite our paper:

@journal{tgsSaltBodiesSegmentation2019,
  title={Semi-Supervised Segmentation of Salt Bodies in Seismic Images using an Ensemble of Convolutional Neural Networks},
  author={Babakhin, Yauhen, and Sanakoyeu, Artsiom, and Kitamura, Hirotoshi},
  journal={German Conference on Pattern Recognition (GCPR)},
  year={2019}
}