Road Segmentation — Computational Intelligence Laboratory Project 2018

This repository summarises our findings and models we used for the task of road segmentation from RGB aerial/satellite images.

All questions should be directed towards the authors in case discrepancies are found or our README is incomplete in instructions.

Setup

The following steps need to be followed to train our networks.

Download Necessary Dataset

Download the training and testing data from the Kaggle page and unzip them into data/.

Installing Dependencies

The following dependencies needed to be fulfilled:

• A fixed Keras-2.2.0
• Tensorflow-1.7.0 or newer
• imgaug-0.2.5

We make brief notes of installation instructions and make references to official installation instructions when possible.

Installing Keras-2.2.0

Checkout branch cil-road-segmentation-2018 and follow normal installation instructions provided by official Keras.

Installing Tensorflow-1.7.0

Install Tensorflow with GPU support. This is required as we make heavy use of NCHW and have not verified NHWC to work.

Installing imgaug-0.2.5

Follow the installation instructions given by imgaug.

Environment Setup

We further require the user to create a configuration file for keras at ~/.keras/keras.json with the following content:

{
    "floatx": "float32",
    "epsilon": 1e-07,
    "backend": "tensorflow",
    "image_data_format": "channels_first"
}

Running Networks

We have the following selection of networks for quick training:

• ResNet18 (resnet18)
• SegNet (segnet)
• RedNet50 (rednet50)

All networks are called with their own train-*.py file for training and predict-*.py file for generating predictions. Their source code can be found in the respective imports.

ResNet18 generates a ready-to-use .csv-file. SegNet and RedNet50 require the use of data/mask_to_submission.py to convert predicted masks to a .csv-file.

Report

The report including source files is in ./report. We publish this work using the Attribution-NonCommercial-ShareAlike (CC BY-NC-SA) license.

Code License

See LICENSE.