Auto Assess Building Damage Based on Siamese Neural Network

This is a project for ECEN649 Pattern Recognition

Docker Environment:

We use Docker to manage our training environment. To start, you need to install the docker at first. Then you can use the following command to download the docker image we uploaded to the Docker Registry:

docker pull cloudtom/ecen649-xview2:v7-plus

Then use:

docker run -it cloudtom/ecen649-xview2:v7-plus

to start the environment. You may need to add -v /path/to/the/code/on/your/machine:/path/in/docker/container to mount the code repo in the Docker container

Example: docker run -it -v /path/to/the/code/on/your/machine:/path/in/docker/container cloudtom/ecen649-xview2:v7-plus

Training:

You need to visit the xView2 challenge official website https://www.xview2.org to download the dataset. And unzip the dataset into the same folder with the code.

Run create_masks.py at first, then run python train_154loc.py for training SENet154 or python train_50loc.py for training ResNet50

You may need to create a wandb.ai account to track the training process.

Predict:

To obtain prediction image on test set, run python pred154_loc.py for prediction with SENet154, run python pred50_loc.py for prediction with ResNet50

Compute F1 score of your result:

Please run F1_score.ipynb, you may need to change the image path. it will output the F1 score compared with the ground truth.

References:

This code is inspired by gihub website: https://github.com/DIUx-xView/xView2_first_place (MIT License)