ML_RoadSegmentation

README for the Machine Learning CS-433: Project 2 - Road Segmentation

Group members:

• Diego Fiori
• Paolo Colusso
• Valerio Volpe

CrowdAI team name: LaVolpeilFioreEilColosso

Architecture

The files created and the functions developed are presented in the following sections:

• Helpers
• Preprocessing
• Regression
• Neural Nets
• Post Processing
• Submission

Helpers

helpers_img.py

Contains the functions to load and read the data, perform basic procsseing of the images, compute the F1 score, and create the submission.

Preprocessing

preprocessing.py

Contains the function to pre-process the images. A series of functions are created to:

• extend the dataset by means of rotation and flip
• extend the borders of the image
• apply filters on the images
• add channels to the image
• extract the features as mean and variance of the channels
• take features of the polynomials

dataset.py

Class used to read the set of images.

Logistic and Ridge Regression

helpers_regression.py

Tools to perform regression with cross validation.

Contains the function used to:

• split the data into train and test set
• call the preprocessing functions
• perform regression
• call the post-processing functions

Cross_Validation_regression.ipynb

• performs regularised logistic regression with cross-validation
• performs ridge regression with cross-validation

Neural Nets

NeuralNets.py: contains the classes fot the Simple Net, the U-Net and the Deep Net.

Bagging_Net.py: contains the functions used to run the bootstrap-like neural net.

The following notebooks can be used to define and run the models:

• Net with bootstrapping: Bagging_Net.ipynb
• U-Net: U-Net.ipynb,
• Deep Net: RUN.ipynb

training_nn.py: contains the functions to train neural networks.

Models: folder containing the the models created.

Postprocessing

Post_processing.py

Contains the functions which perform post-processing operations on the predictions obtained for the images from either of the models mentioned above.

Submission

mask_to_submission.py submission.py submission_to_mask.py

CrowdAI results

Username: Paolo Colusso

Submission ID Number: 25160

References

• Statistical learning: James, Witten, Hastie, Tibshirani, Introduction to Statistical Learning, see details.

• Image processing: Burger, Burge, Digital Image Processing. An Algorithmic Introduction Using Java, see details.

• U-Net: Ronneberger, O., Fischer, P., and Brox, T., U-Net: Convolutional Networks for Biomedical Image Segmentation, 2015.

• Neural nets: EPFL course available here.