Team Member (CWX):

• Zewei Xu
• Haoxuan Wang
• Ganyuan Cao

We implemented a model based on the LinkNet architecture to classify the road segments with the satellite images from Google Maps. We evaluate the performance of LinkNet, U-Net and their variants. We finally decided to develop our model on the top of LinkNet architecture.

In addition, we implemented U-Net architecture (and its variants: U-Net++) as a comparison to show the performance of our model.

/data: The folder that contains the training and the testing images, it is also used to save the prediction masks.

/log: The folder that contains the evaluation results of different models.

Losses.py: Loss Functions including BCE Loss, Dice Loss, Focal Loss and a mix loss combined by Focal Loss and Dice Loss.

dataset.py: Load images and perform data augment on images.

Helper.py: Some helper functions and image transformation.

mask_to_submission.py: Convert the prediction results into a submission file.

model.py: Model implementation.

• Unet
• LinkNet
• LinkNet1
• DinkNet
• UnetPP
• LinkNetPP
• DoubleUnet

run.py: It creates a model and train it, or loads a pre-trained model, and it makes predictions on the test set.

submission_to_mask.py: Convert the submission file into prediction images.

train.py: It trains the model.

utile.py: It contains some utile functions, such as:

• get_loader: Create the DataLoader class
• f1: Compute F1 score
• accuracy: Compute F1 score, Precision and Recall
• create_different_prospective: Apply transformations to the image and return different prospectives
• combine_different_prospective: Combine predictions of different prospectives
• img_crop: Crop the image into patches of a given size
• combine_img: Combine cropped patches into the original image

visualization.ipynb: A visualization of performance of different models.

• Python 3.8.5
• Jupyter Notebook 6.1.4

• Numpy 1.20.3
• PyTorch 1.8.2
• PIL 8.2.0
• albumentations 1.1.0
• tqdm 4.59.0
• matplotlib 3.3.4
• torchsummary
• pandas 1.2.4
• torchvision

To install all the required libraries, you can run the following command from the terminal.

pip install -r requirement

Our pre-trained models can be found here:

https://drive.google.com/drive/folders/1VPxJjSSlY1VeJWWgAL2eXIICsWvUjpjo?usp=sharing