ways_to_improve.md

Imrpovements model

Currently, I'm not use the loss function as suggested in the original paper. So let's try with the loss that the author in paper [1] has suggest

High priority with the idea from [1] at reference [28]

• Reread [1] paper, notice at references:
  • 27 - A new deep learning architecture for detection of long linear infrastructure
  • 28 - A novel focal phi loss for power line segmentation with auxiliary classifier u-net

Implement and train the model with Jaccard Loss

• Research on IoU loss and the way people make it become differentiable when apply as the loss function.
• Links to read:
  • StackOverflow discussion - Why people does not use the IoU for training model
  • Paper
  • Pytorch's Implementation

Papers

• [1] PLGAN: Generative Adversarial Networks for Power-Line Segmentation in Aerial Images
• [2] DUFormer : A Novel Architecture for Power Line Segmentation of Aerial Images
• [3] Automatic High Resolution Wire Segmentation and Removal