This repository provides a new and more fine-grained annotated distorted document training dataset called DocDewarpHV.
This dataset contains about 110K distorted document images in Chinese and English. The number of Chinese and English documents is close to 1:1. The resolution of each image is 512*512. The source scanned images come from cddod, CDLA, M6Doc and PubLayNet. Compared with Doc3D, in addition to 3D world coordinates, UV map, 2D backward map (grid coordinates), we also provide horizontal and vertical line annotations that are consistent with the distortion trend of the input image.
DocDewarpHV/
alb_h/
cddod_1/
1-0_ann0001.png
1-1_ann0001.png
...
CDLA_1/
M6Doc_test_1/
publaynet_train_1/
...
alb_v/
cddod_1/
1-0_ann0001.png
1-1_ann0001.png
...
CDLA_1/
M6Doc_test_1/
publaynet_train_1/
...
bm/
cddod_1/
1-0_ann0001.mat
1-1_ann0001.mat
...
CDLA_1/
M6Doc_test_1/
publaynet_train_1/
...
uvmat/
cddod_1/
1-0_ann0001.mat
1-1_ann0001.mat
...
CDLA_1/
M6Doc_test_1/
publaynet_train_1/
...
warp_img/
cddod_1/
1-00001.png
1-10001.png
...
CDLA_1/
M6Doc_test_1/
publaynet_train_1/
...
wc/
cddod_1/
1-0_ann0001.exr
1-1_ann0001.exr
...
CDLA_1/
M6Doc_test_1/
publaynet_train_1/
...
DocDewarpHV.txt
You can download the entire DocDewarpHV dataset from Baidu Netdisk (size: ~600GB). We will upload it soon.
You can directly execute the python file doc_dewarp_hv_read.py as follows.
Remember to modify the dataset path in the main function.
This code is also applicable to reading data when training your own rectification model.
python loader/doc_dewarp_hv_read.py
The DocDewarpHV dataset should be used under CC BY-NC-ND 4.0 License for non-commercial research purposes.
If you have any questions about this dataset, you can always contact hengli.lh@outlook.com
Thanks to Doc3D, the code for this DocDewarpHV data synthesis is based on it. We also thanks to cddod, CDLA, M6Doc and PubLayNet for their outstanding work in open-sourcing the original document images.