Skip to content

Latest commit

 

History

History
42 lines (33 loc) · 1.55 KB

README.md

File metadata and controls

42 lines (33 loc) · 1.55 KB
Raw

Improved_Control_Points

Quick Start

This is an improved document dewarping method based on Control Points.

clone

git clone https://github.com/NikLi66/Improved_Control_Points icp
cd icp

distributed training

bash ddp_train.sh

Note: you could change the output path, data path and the number of gpus if necessary.

test

bash test.sh

evaluate

python eval.py

Contributions

  • Rewritting most part of the original codes to make it more readable.
  • Adding CBAM modules.
  • Adding Coord Conv modules.
  • Adding some data augmentations widely used in dewarp task.
  • Optimizing the loss function by adding a weight mask.

Training on your own dataset

I strongly recommend you slightly change the codes in dataset/dataloader.py to make it works for your onw dataset. You only need to change the codes in getitem and init functions I suppose. The data format used in this project is following Control Points. Please find more details in that repository if you like.

Acknowledge

  • Document Dewarping with Control Points paper github
  • CBAM: Convolutional Block Attention Module paper
  • An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution paper github