yutingyao/steel_defect_detection

Kaggle 2019 steel defect detection contest, final project for course ECE285

ECE 285 Final Project (D): Severstal Steel Defect Detection

Authors: Chi-Hsin Lo, Subrato Chakravorty, Utkrisht Rajkumar

Description

This is project D: Severstal Steel Defect Detection using semantic segmentation by team 3Nature_2Journal_8GPU.

The dataset for this project can be found here. To run with our default path setup, create a new folder called "inputs" in the same directory as the one containing this cloned respository. Download and extract the dataset inside "inputs." The kaggle submission kernel can be found here. It contains the code to upload submission.csv files and submit to Kaggle. Inside the kernel, load the dataset called ece285_submissions to access all generated submission files. The submissions can be directly downloaded from kaggle here.

Requirements

All the package requirements can be found in requirements.txt. To install the requirements:

git clone https://github.com/ucrajkumar/ece285
cd ece285
pip install -r requirements.txt

If working in a conda environment, use:

conda install --file requirements.txt

The required file structure to run with our default paths are as follows: File Structure

• Current Directory
  • Inputs
    • Downloaded data from Kaggle
  • ece285 (this repo)

Code organization

file name	Description of file
Demo.ipynb	Qualitatively evaluate all models on validation and test set images
Figures.ipynb	Recreate figures from report
Train.ipynb	Train u-net, u-net with residual encoder, and u-net with inverted residual encoder
deeplabv3.ipynb	Train deeplabv3+
Inference.ipynb	Predict on test set and export results
data_gen.py	Data generator for loading train, validation, and test images
model.py	Code for building all 4 models
utils.py	Accuracy and loss metrics, conversion from mask to RLE and vice-versa, and post-processing
train_idx.npy	indices for training set
val_idx.npy	indices for validation set

---

folder name	Description of file
ex_images	Example validation and test images for Demo.ipynb
history	Training history for each model
models	Trained model files
submissions	Exported results on test set

*Note: The submission for deeplab was too large to upload onto github. However, it is in ece285_submissions folder in Kaggle.