This repo is the offical implementation for the paper "U-Net based skeletonization and bag of tricks" in the Pixel SkelNetOn Challenge in the "Deep Learning for Geometric Computing" workshop at ICCV 2021

Our solution includes:

1. The modification of U-Net architecture using the attention mechanism.
2. Auxiliary task learning for a more effective training process.
3. Tricks for improving the skeletonization model's performance.

1. Clone https://github.com/namdvt/skeletonization
2. python 3, cv2, pytorch > 1.2

1. Place the SkelNetOn dataset into 'dataset/Starting Kit Pixel/'
2. Check config in "configs/unet_att.yaml"
3. Run the command to train test the model

python main.py