chargrid

Chargrid

This code repository contains the implementations of the paper Chargrid: Towards Understanding 2D Documents (EMNLP, 2018).

Preparing Dataset

• Wildreceipt: Original dataset can be downloaded from MMOCR. Formatted training datalist and vocab used can be found in demo/text_ie/datalist/wildreceipt.
• PubLayNet: PubLayNet is a large dataset of document images, of which the layout is annotated with both bounding boxes and polygonal segmentations. To perform multimodal layout analysis task, we also need to extract the annotations in the character granularity, in addition to layout component granularity. We provide demo datalist examples in demo/text_ie/datalist/PubLayNet, and one can get the above annotation through this instruction.

Train From Scratch

Temporarily, we release pretrained model on Wildreceipt (for visually information extraction task) dataset and PubLayNet (for layout reocognition task) dataset using OCR ground-truth.

If you want to re-implement the model's performance from scratch, please following these steps:

1.Firstly, download original dataset and unzip it to an arbitrary folder.

2.Secondly,

• For Wildreceipt information extraction task, modify the paths in demo/text_ie/chargrid/configs/wildreceipt_chargrid.py including images path, datalists path, work space, etc.
• For PubLayNet layout analysis task, modify the paths in demo/text_ie/chargrid/configs/publaynet_chargrid.py including images path, datalists path, annotation files path, work space, etc.

3.Thirdly, directly run demo/text_ie/chargrid/dist_train_wildreceipt.sh or demo/text_ie/chargrid/dist_train_publaynet.sh.

We provide the implementation of online validation

Test

Given the trained model, directly run demo/text_ie/chargrid/test_wildreceipt.sh or demo/text_ie/chargrid/test_publaynet.sh to inference it.

To visualize generated chargrid map and PubLayNet detection result, you can modify the paths (vis_dir in demo/text_ie/chargrid/test_chargrid.py and vis_save_dir in test pipeline dict of files in demo/text_ie/chargrid/configs/) in testing and config scripts, then start testing:

python test_chargrid.py 

Some visualization of generated chargrid maps and detection results are shown:

• WildReceipt

• PubLayNet

Trained Model Download

All of the models are re-implemented and well trained based on the opensourced framework mmdetection. So, the results might be slightly different from reported results.

Results on various datasets and trained models can be download as follows:

Dataset	Input	Task	Test Scale	metric	result	Links
Wildreceipt	Image	Visually Information Extraction	(512, 512)	F1-score	67.10	config, pth (Access Code: P69m)
Wildreceipt	Image+Chargrid	Visually Information Extraction	(512, 512)	F1-score	81.02	config, pth (Access Code: RL1o)
PubLayNet	Image	Layout Recognition	(1300, 800)	mAP	71.9	config, pth (Access Code: Qc71)
PubLayNet	Image+Chargrid	Layout Recognition	(1300, 800)	mAP	71.8	config, pth (Access Code: 6ng0)

The PubLayNet results are report based on a sampled sub-set (1000) of the original dataset.

Citation

@inproceedings{DBLP:conf/emnlp/KattiRGBBHF18,
  author    = {Anoop R. Katti and
               Christian Reisswig and
               Cordula Guder and
               Sebastian Brarda and
               Steffen Bickel and
               Johannes H{\"{o}}hne and
               Jean Baptiste Faddoul},
  title     = {Chargrid: Towards Understanding 2D Documents},
  booktitle = {EMNLP},
  pages     = {4459--4469},
  year      = {2018},
}

License

This project is released under the Apache 2.0 license

Copyright

If there is any suggestion and problem, please feel free to contact the author with qiaoliang6@hikvision.com.