U-Net based skeletonization and bag of tricks

Skeletonization is a process focused on providing a compact and simple representation of an object by extracting the skeleton pixels from the given shape in a binary image. This method has been widely applied in various image processing and computer vision applications. In addition to traditional approaches which are not robust and provide low accuracy results, many efforts have been made for creating deep learning based methods to overcome these disadvantages. However, skeletonization is still a new topic in the deep learning world. In this paper, we propose our solution for the Pixel SkelNetOn challenge in the third edition of the "Deep Learning for Geometric Computing" workshop at ICCV 2021, which includes (1) modification of UNet architecture using the attention mechanism, (2) implementation of auxiliary task learning for a more effective training process and (3) application of several tricks for improving the skeletonization model's performance. Our method achieved 0.8000 on the Pixel SkelNetOn validation set and second place in the leaderboard.