PAN-reimplement

This is an unofficial pytorch re-implementation of "Efficient and Accurate Arbitrary-Shaped Text Detection with Pixel Aggregation Network"(19'ICCV,PAN/PSENet-v2) and some personal modifications.

• Thanks to the help of the official PSENet implementation[https://github.com/whai362/PSENet] and PAN.pytorch[https://github.com/WenmuZhou/PAN.pytorch]. This repo is mainly based on the PSENet repo and followed a similar framework.
• While PSENet is implemented in Python2, this repo uses Python3 to implement both PSENet and PAN. Compared to the PAN.pytorch repo, this repo fixes some bugs and achieves a higher f-score(but is still lower than the original paper by ~0.5%, sigh)
• The author of PSENet and PAN has already released the official implementation of PAN[https://github.com/whai362/pan_pp.pytorch]. Since my re-implementation can not perfectly reproduce the results, I suggest that referring to the official repo might be a better choice.

Packages required

• pytorch 1.1.0 & torchvision
• Polygon3
• pyclipper
• pybind11(to compile the cpp version pse algorithm)

How to train PSE

• train on icdar
  CUDA_VISIBLE_DEVICES=0,1,2,3 python train_icdar15.py --backbone res18/res50 --lr YOUR.LR --resume YOUR.CHECKPOINT_FILE
• train on ctw
  CUDA_VISIBLE_DEVICES=0,1,2,3 python train_ctw1500.py --backbone res18/res50 --lr YOUR.LR --resume YOUR.CHECKPOINT_FILE

How to train PAN

• train on icdar
  CUDA_VISIBLE_DEVICES=0,1,2,3 python train_icdar15_PAN{_v2}.py --backbone res18/res50 --lr YOUR.LR --resume YOUR.CHECKPOINT_FILE
• train on ctw
  CUDA_VISIBLE_DEVICES=0,1,2,3 python train_ctw1500_v2.py --backbone res18/res50 --lr YOUR.LR --resume YOUR.CHECKPOINT_FILE
• Other training scripts are variants which have external extension like top/bottom boundary prediction. You can check the codes for detailed information

Experimental Results

(TODO)

Future work & Note

• This version is just a preliminary manuscript and has not been organized well. Sorry for organizing the codes in such a messy way, I'll update a cleaner version as soon as possible.
• The detailed performance of PAN/PSE (this repo's performance has a slightly drop compared to the original PAN paper)
• and a guide to the training process will also be uploaded soon. (done)
• Detailed information of my personal modifications(top/bottom boundary prediction) will also be released in the future.
• If this repo really helps you, please star it, thx!