segme-net

Road Segmentation Project for Computational Intelligence Lab 2019

Group: Seggy Roady Projecty

Martin Blapp, Laurin Paech, Nihat Isik, Qais El Okaili

Department of Computer Science, ETH Zürich, Switzerland

github link: https://github.com/laurinpaech/segme-net

ATTENTION!!! We had to delete train_600 (the 600 additional training images). If you want to have the additional data, please download it from the github repo and insert it according to the folder structure in directory data.

Folder structure

├──  data
|   |
|   |── train                       - all 100 training images given for the project
|   |
|   |── valid                       - validation images generated from chicago dataset
|   |   
│   |── test                        - all test images given for the project
|   |
|   └── train_600                   - training images + additional 600 generated images
│
│
├── model                               - this folder contains any model of our project.
|   |
|   |── encoder_decoder.py                  - keras implementation of simple encoder decoder
|   |
|   |── segnet.py                           - keras implementation of SegNet
|   |
|   |── stacked_unet.py                     - keras implementation of stacked unet
|   |
|   |── stacked_unet_leaky.py               - keras implementation of stacked unet + leaky relu
|   |
|   |── stacked_unet_leaky_wavelet.py       - keras implementation of stacked unet + leaky relu + wavelets
|   |
│   └── stacked_unet_leaky_wavelet_2.py     - keras implementation of stacked unet + leaky relu + wavelets
│
│
├── main.py                        - main that is responsible for the whole pipeline
│ 
│  
├── data _loader
|   | 
│   └── data.py                 - data generator that is responsible for all data handling
│ 
└── utils
     |
     ├── overlay_generator
     |   |
     |   ├── normal_img             - contains original images for that we want to generate overlays of predictions
     |   |
     |   ├── submit_img             - contains images that are used as overlay for normal_img
     |   |
     |   └── overlay_generator.py   - generates 
     |
     ├── alpha_testing.py           - used for testing best cut-off value for our predictions
     |
     ├── custom_losses.py           - custom loss functions
     |
     ├── custom_layers.py           - custom layers for segnet
     |
     ├── hough_transform.py         - Probabilistic Hough Transform to fill gaps in results
     |
     ├── mask_to_submission.py
     |
     ├── submission_to_mask.py
     |
     └── image_generation.py        - used for generation additional images as data

Getting started

Use python 3.6 and run the following command:

pip install -r requirements.txt

Stacked U-Net, using Keras

The architecture was inspired by U-Net: Convolutional Networks for Biomedical Image Segmentation.

How to use

Running a model

first load env on leonhard

module load gcc/4.8.5 python_gpu/3.6.4 hdf5 eth_proxy
module load cudnn/7.0

2.. then run (estimated runtime 8 hours on a "GeForce GTX 1080 Ti")

python main.py --desc "reproducing_best_result" --epochs 1000 --rotation 360 --width_shift_range 50 --height_shift_range 50 --shear_range 10 --zoom_range 0.1 --horizontal_flip --fill_mode "reflect" --nr_of_stacks 2 --resize --ensemble

result will be located in data/roadseg/submit_output/reproducing_best_result

How to create submission

run the model with python main.py with appropriate flags
predictions are in data/submit_output/
run mask_to_submission.py on output folder
next_submission.csv in folder can now be uploaded to kaggle

Reproducing Kaggle results

run the following command:

python main.py --desc "stacked_unet_2stack" --epochs 1000 --rotation 360 --width_shift_range 0 --height_shift_range 0 --shear_range 0 \
--zoom_range 0 --horizontal_flip --fill_mode "reflect" --nr_of_stacks 2 --resize --ensemble

This runs the Stacked U-Net on the 100 training images and saves the result in data/submit_output/stacked_unet_2stack/.

NOTE: if the accuracy doesn't increase within the first 30 epochs, run again.