Beyond the Pixel-Wise Loss for Topology-Aware Delineation

While many deep learning based frameworks are proposed to solve the delineation problem in images, the global topology can not be guarenteed in most cases. The authors designed a topology-aware U-Net in the paper: Beyond the Pixel-Wise Loss for Topology-Aware Delineation.

In this paper, we achieved the following goals:

• Implemented the algorithm in PyTorch

• Reproduced the results on Massachusetts Roads Dataset

Requirements

• PyTorch 1.0
• >= Python 3
• Visdom

Source Code

Code / Report / Slide