cs-433 / thehappywhales_road_segmentation

EPFL_Road_Segmentation created by GitHub Classroom

Road Segmentation /AIcrowd challenge

ml-project-2-the-happy-whales2 created by GitHub Classroom

The goal of the project is to segment roads in satellite/aerial images acquired from GoogleMaps. The training dataset consists in 100 RGB images of size 400x400x3 and their corresponding groundtruth in black and white . To detect roads, three classifiers are created in this project and their efficiency compared. In this project, patches of size 16x16 are classified: if a road is detected in the patch the classifier ouputs 1 and if not it ouputs 0. An example of an aerial image used for training and its mask is shown below:

The set of training and test images is available on this page: [EPFL ML Road Segmentation challenge](https://www.aicrowd.com/challenges/epfl-ml-road-segmentation).

Prerequisites

This project has been realized using:

• python3.9
• numpy
• segmentation_models
• patchify
• matplotlib
• keras
• tensorfow= 2.7.0
• openCV
• scipy
• pandas

Structure of files

The folder MLprojet has can be run on Google Colab to use the GPU or one's computer. Make sure that the files of ML_project are identical to the file structure shown below, before trying to run the code one the computer.
It contains several sub-folders:

ML project

|-- data
|        |-- training: original training dataset
|        |          |-- images (100 satellite images)
|        |          |        |--satImage_001.png
|        |          |        |-- ...
|        |          |-- groundtruth (100 groundtruth images)
|        |          |        |--satImage_001.png
|        |          |        |-- ...
|        |-- augmented_training: created augmented dataset for the training
|        |          |-- images (1000 satellite images)
|        |          |        |--satImage_001.png
|        |          |        |-- ...
|        |          |-- groundtruth (1000 groundtruth images)
|        |          |        |--satImage_001.png
|        |          |        |-- ...
|        |-- test_set: original testing dataset
|        |          |-- images (50 satellite images)
|        |          |        |--test_1.png
|        |          |        |-- ...
|-- src: script files
|        |-- train.py: train the different models implemented, first lines have to be modified accoriding to the desired model
|        |-- run.py: create predictions for the testing dataset and create and save the submission file
|        |-- data_augmentation.py: create the augmented training dataset
|        |-- models: files with functions for the definition of the models
|              |-- cnn_model.py: for the CNN model
|              |-- unet.py: for the U-Net model
|        |-- utilisites: files with necessary functions
|              |-- data_preprocessing.py: functions for the preprocessing on the images before training or testing
|              |-- helpers_train.py: functions for the training of the models
|              |-- helpers_run.py: functions for the test of the models and the submission
|-- saved_models: this folder is not in the github folder as it was to large. One can download it from the Google drive folder.
|        |-- cnn_noaug.h5: final CNN model trained on original dataset
|        |-- cnn_aug.h5: final CNN model trained on augmented dataset
|        |-- unet_noaug.hdf5: final U-Net model trained on original dataset
|        |-- unet_aug.hdf5: final U-Net model trained on augmented dataset
|        |-- resnet34_noaug.hdf5: resnet34 model for the final Multinet model trained on original dataset
|        |-- seresnet34_noaug.hdf5: seresnet34 model for the final Multinet model trained on original dataset
|        |-- resnet50_noaug.hdf5: resnet50 model for the final Multinet model trained on original dataset
|        |-- resnet34_aug.hdf5: resnet34 model for the final Multinet model trained on original dataset
|        |-- seresnet34_aug.hdf5: seresnet34 model for the final Multinet model trained on original dataset
|        |-- resnet50_aug.hdf5: resnet50 model for the final Multinet model trained on augmented dataset
|-- prediction_unet: predicted masks of the best model (Multi-Resnet trained on augmented dataset, corresponding to file submission_final.csv)
|        |-- pred_1_unet.png
|        |-- ...
|-- submissions: submission files for AIcrowd, only 6 of them are stored here, the ones corresponding to the final results presented in the report
|        |-- submission_cnn_noaug.csv: final submission for the CNN model on original dataset
|        |-- submission_cnn_aug.csv: final submission for the CNN model on original dataset
|        |-- submission_unet_noaug.csv: final submission for the U-Net model on original dataset
|        |-- submission_unet_aug.csv: final submission for the U-Net model on augmented dataset
|        |-- submission_multinet_noaug.csv: final submission for the Multi-resnet model on original dataset
|        |-- submission_multinet_aug.csv: final submission for the Multi-resnet model on augmented dataset
|        |-- submission_final.csv: final submission for the final model, that gave the best results: Multi-resnet model on augmented dataset

Usage

One can access the saved models in the google drive: https://drive.google.com/drive/folders/1Yt2N6YO7XZkUdVjKsYC2w7NswY3phRAQ?usp=sharing
One can either run the model from google colab with the file train.ipynb and run.ipynb or from
one's computer with train.py and run.py

How to train the model

In train.py/ipynb:

• Choose patches with PATCHES = 256 or 128
• TRAINING = 1: training with no validation set, TRAINING = 0: training with validation set
• TEST_SIZE = choose test size <1
• Choose EPOCHS
• Choose BATCH_SIZE
• Choose model with MODEL = "UNET" or "RESNET34" or "SERESNET34" or "RESNET50" or "CNN"

Run train.py with : python3.9 src/train.py

How to run the model

In run.py/ipynb:

• Choose patch size with PATCHES = 256 or 128
• Choose number of models predicted. MODEL_NR = 3: ensemble prediction of 3 models, MODEL_NR = 1: prediction on 1 model
• If model is CNN, CNN = True

Run run.py with :
python3.9 src/run.py

Contributors

• Charlotte Sertic
• Estelle Chabanel
• Servane Lunven