README.md

Note: This README is an early work-in-progress. The body still needs to be written and images need to be resized.

Implementing U-Net using the TensorFlow Estimator API

This repository contains scripts for building, training, and generating predictions with our own implementation of the U-Net architecture. The network was implemented in Python using TensorFlow's Estimator API. Our goal was to reproduce the U-Net architecture described in [1], then train the network on data provided by the Carvana Image Masking Challenge hosted on Kaggle [2]. Our implementation produced masks with average pixel accuracy of 0.9959 +/- 0.0003 over a 6-fold cross-validation set.

Example masks produced by a single network trained on one fold of the cross-validation set can be seen below.

Description of the original U-Net architecture

U-Net is a deep, fully convolutional neural network architecture proposed for biomedical image segmentation. A visual representation of the network, as shown in the original publication [1], can be found below.

Differences between our implementation and the original architecture

Summaries of each file

Running the program

References

1. U-Net: Convolutional Networks for Biomedical Image Segmentation
   Olaf Ronneberger, Philipp Fischer, Thomas Brox.
   [link]. arXiv:1505.04597, 2015.

2. Carvana Image Masking Challenge
   Carvana LLC.
   [link]. Kaggle, 2017.