Road Segmentation from Aerial images

This is the projects source code of Robin Bader, Francesco Saverio Varini and Jakob Beckmann which is part of a Kaggle competition that is available under this link this link.

Quick start

To quickly get started and train and predict the final submission obtained in the Kaggle competition executed following code in the command line assuming that <project-root> refers to this projects root directory.

cd <project-root>
python ./src/run.py -t -m u_net
### Wait for training to finish
python ./src/run.py -p -m u_net
### File csv output will be available under: ./trained_models/u_net/<last-entry>/submission_u_net_<timestamp>.csv

In case this does not run, please read further details below.

Project structure

Following list describes the directories and their use:

Folder	Description
assets	This folder contains all the data needed to train and test the model.
report	Contains the code to generate the projects report.
src	This folder contains the projects source code.
trained_model	This folder is automatically generated during the training of the model and will contain the projects checkpoints and the final submission.csv
<project-root>	Contains this README file. All code needs to be executed based on this path. Additionally holds the report.pdf and the declaration_of_originality.pdf

Prerequisites

In order to run this project following requirements need to be met.

Environment Setup

This project is based on Python 3.6 and Tensorflow (tested with version 1.7 and 1.8). Therefore the environment needs to be set up with following packages:

numpy
matplotlib
keras (dependent on tensorflow and h5py)
tensorflow (tested with version 1.8)
pandas
Pillow for PIL

Training the model

The score on Kaggle was achieved by training the model until the project automatically finished improving the validation error.

Usage

usage: run.py [-h] [-m {cnn_lr_d,u_net,u_net_dropout}] [-t] [-tr]
              [-d DATA] [-p] [-vis VISUALIZE]

Control program to launch all actions related to this project.

optional arguments:
  -h, --help            show this help message and exit
  -m {cnn_lr_d,u_net,u_net_dropout}, --model {cnn_lr_d,u_net,u_net_dropout}
                        the CNN model to be used, defaults to u_net_dropout
  -t, --train           train the given CNN
  -tr, --train_resume   continue training the given CNN
  -d DATA, --data DATA  path to the data to use (prediction)
  -p, --predict         predict on a test set given the CNN
  -vis VISUALIZE, --visualize VISUALIZE
                        visualize prediction of an image given its id

Executing the training

To start the training execute following command:

python <project-root>/src/run.py -t -m u_net

Important: The project needs to be executed out of the project-root folder. E.g. the current directory must be the project-root!

Executing the prediction

To start the prediction, execute following command:

python <project-root>/src/run.py -p -m u_net

The output is generated under:

<project-root>/trained_models/u_net/<start time training>/submission_u_net_<timestamp>.csv

Visualizing the predictions

In order to visualize the predictions, please first make sure to have first generated the submission.csv. Then, just digit:

python <project-root>/src/run.py -vis <id-number-test-image> -m u_net

bastianmorath/cil_project