This repository contains ipython implementation of U-net network for car segmentation in 2D based on grayscale or RGB color images. The training and test dataset are from carvana Kaggel challenge (144 images and 144 ground-truth masks). The model has been tested on Linux and Windows 10. Using this example, good result can be achived with in a few minutes depending on number/type of the GPUs.

This implementation is based on Keras with Tensorflow backend. It has been tested on Python 3.6, Anaconda (4.7). The uplaoded code requires a GPU but if you do not have one, simply uninstall tensorflow-gpu (pip uninstall tensorflow-gpu), then install regular tensorflow (pip install tensorflow) and the code will work.

This can serve as an example for learning to:

• Train a network using your own images

• Learn the structure of Unet netwrok

• Test and fine tune a network to improve the segmentation task

• Perform 2D segmentation with other datasets

• Unzip data folder
• Create a virtual environment with Anaconda: conda create -n DL_example python=3.6 (Python 3 is compatible with packages)
• Activate the environment: activate DL_example
• Change directory to where you cloned the current files: cd ./where files are downloaded
• Install required libraries: pip install -r requirements.txt

To test the notebook version, type the following in the command line: jupyter notebook This will open the notebook which will allow you to run either the grayscale (car_segmentation_Unet_grayscale.ipynb) or RGB (car_segmentation_Unet_rgb.ipynb) version of the code. You should be able to run each cell and get the result.

Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional networks for biomedical image segmentation." In International Conference on Medical image computing and computer-assisted intervention, pp. 234-241. Springer, Cham, 2015.

The code comes "AS IS" with no warranty of any kind. It can be used for any educational and research purpose. Feel free to modify and/or redistribute.