cs-433/cs-433-project-2-still-no-idea

cs-433-project-2-still-no-idea created by GitHub Classroom

Project Road Segmentation

This is the code and report for the Road Segmentation project of the CS-433 class.

Report

The report can be found in Report.pdf

Required libraries

To run the notebooks in local you will need the following libraries :

• matplotlib
• pyTorch v1.5.1 and torchvision v0.6.1 using conda install pytorch==1.5.1 torchvision==0.6.1 cudatoolkit=10.2 -c pytorch
• scikit-image
• opencv-python

In run.py you can update the line 26 root_dir = "" with the path to this directory if needed. Then you can execute run.py in local, which will give a file final_submission.csv.

Code organization

The code is split into two main parts:

• The UNet architecture is contained in the Unet.ipynb file
• The CNN architecture is contained in the Road_segmentation.ipynb or run.py file

Both file were run on Google colab and contains a variable root_dir which indicates the directory with the data and the pretrained model. This root directory must have the following structure :

• images/
  • all images for the training set
• groundtruth/
  • all groundtruth for the training set
• test_set_images/
  • test_i/ : for i from 1 to 50
    • test_i.png
• results/ : the folder to save the prediction
• convmodel100.pth : pretrained CNN with 100 epochs
• modelUnet.pth : pretrained UNet with 100 epochs

mask_to_submission.py transforms the submission found in a results directory into a submission.csv file

Alternatively, you can run run.py with the complete repo. The structure for the CNN is already there.

Pretrained models

convmodel100.pth contains the pretrained CNN with 100 epochs modelUnet.txt contains a link to get the pretrained UNet with 100 epochs. The file is too big for GitHub

Results

results_cnn/ contains the predictions made by the CNN
results_postproc/ contains the postprocessed predictions made by the CNN
results_unet/ contains the predictions made by the UNet

AIcrowd

The best submission is Submission #109779

Authors

• Simon Wicky [email protected]
• Gaultier Lonfat [email protected]
• Joachim Dunant [email protected]