zzm422 / dewarpnet

Code for the paper "DewarpNet: Single-Image Document Unwarping With Stacked 3D and 2D Regression Networks" (ICCV '19)

DewarpNet

This repository contains the codes for DewarpNet training.

Recent Updates

• [May, 2020] Added evaluation images and an important note about Matlab SSIM.

Training

• Prepare Data: train.txt & val.txt. Contents should be like:

1/824_8-cp_Page_0503-7Ns0001
1/824_1-cp_Page_0504-2Cw0001

• Train Shape Network: python trainwc.py --arch unetnc --data_path ./data/DewarpNet/doc3d/ --batch_size 50 --tboard
• Train Texture Mapping Network: python trainbm.py --arch dnetccnl --img_rows 128 --img_cols 128 --img_norm --n_epoch 250 --batch_size 50 --l_rate 0.0001 --tboard --data_path ./DewarpNet/doc3d

Inference:

• Run: python infer.py --wc_model_path ./eval/models/unetnc_doc3d.pkl --bm_model_path ./eval/models/dnetccnl_doc3d.pkl --show

Evaluation:

• We use the same evaluation code as DocUNet. To reproduce the quantitative results reported in the paper use the images available here.

• [Important note about Matlab version] We noticed that Matlab 2020a uses a different SSIM implementation which gives a better MS-SSIM score (0.5623). Whereas we have used Matlab 2018b. Please compare the scores according to your Matlab version.

Models:

• Pre-trained models are available here. These models are captured prior to end-to-end training, thus won't give you the end-to-end results reported in Table 2 of the paper. Use the images provided above to get the exact numbers as Table 2.

Dataset:

• The doc3D dataset can be downloaded using the scripts here.

More Stuff:

• Demo
• Project Page
• Doc3D Rendering Codes

Citation:

If you use the dataset or this code, please consider citing our work-

@inproceedings{SagnikKeICCV2019, 
Author = {Sagnik Das*, Ke Ma*, Zhixin Shu, Dimitris Samaras, Roy Shilkrot}, 
Booktitle = {Proceedings of International Conference on Computer Vision}, 
Title = {DewarpNet: Single-Image Document Unwarping With Stacked 3D and 2D Regression Networks}, 
Year = {2019}}   

Acknowledgements:

• These codes are heavily structured on pytorch-semseg.