1. Introduction:

KVP10k is a dataset of 10,000 pages annotated for Key-Value Pair (KVP) extraction from business documents, featuring linked rectangles, from value to its matching key, and text labels.
We also provide code for downloading and preparing the dataset, model training, and benchmarking, serving as a foundation for model comparison and evaluation, as detailed in our paper.

The related paper can be found here: here

1.1. Data Source

The data used in this project is downloaded from Hugging Face. You can access and explore the dataset directly through this link.

Please ensure to check the dataset's usage rights and licensing terms before using it in your projects.

2. Downloading and Preparing the dataset for traininig and evaluation

2.1. The KVP10k dataset

The KVP10k dataset contains 10,000 annotated pages, each containing annotation matching for KVP extraction tasks, in the following format.
The annotations correspond to images that have been extracted from PDF documents.

{
  "rectangles": [
    {
      "_id": "48758be2-7508-4f9c-8198-3cbc867d8ce0",
      "attributes": {
        "Linking": {
          "state": "editable",
          "value": "71dae826-d65b-45ae-bcfa-04b9fe754315"
        }
      },
      "color": "rgb(253, 255, 0)",
      "coordinates": [
        {"x": 0.180248, "y": 0.593932},
        {"x": 0.282669, "y": 0.593932},
        {"x": 0.282669, "y": 0.613686},
        {"x": 0.180248, "y": 0.613686}
      ],
      "label": "Text",
    }
  ],
  "image_url": "https://ae.ucr.edu/sites/default/files/2019-05/DPP2005-r15-FINAL.pdf",
  "page_number": 212,
  "hash_name": "0a050bcc3ace2ed3907d22f8dacc5adf0d960d8c5c5a692afd72a17da13b7ee4"
}

• We define unkeyed_value type list: name, date, address, phone, email, website, year, document type, document title, text
• The annotation per each image contains:
  • A list of rectangles described by:
    • _id: the rectangle id
    • coordinates: 4 coordinates representing the box around the key or value (the coordinates are relative to the page size and take values between 0 and 1)
    • label: Text, unvalued_key or one of the values from unkeyed_value type list
    • Linking: the id of the rectangle key (empty if the rectangle is key, unvalued_key or unkeyed_value)
  • Additional metatdata:
    • image_url: the url to download the pdf
    • page_number: the pdf page number
    • hash_name: name of the image

Annotation example

2.2. Preparing the dataset

2.2.1. Prerequisite:

Download tessdata

mkdir /path/to/tessdata
wget --quiet https://github.com/tesseract-ocr/tessdata/raw/4.0.0/eng.traineddata -O /path/to/tessdata/eng.traineddata

2.2.2. Setting up environment for only preparing the dataset :

We used conda version 24.5.0

git clone git@github.com:IBM/KVP10k.git
cd KVP10k/
                    
conda create -y -n KVP10k python=3.11  
conda activate KVP10k   

sudo apt-get install poppler-utils               

pip3 install --only-binary "matplotlib" --only-binary "tesserocr"  -e '.'

2.2.3. Download and prepare dataset:

We provide in this repository a download_dataset.py script that for each annotation json:

• Download the matching pdf following the image_url and extracting the image from page_number
• Run OCR (tesseract) on the extracted image
• Create the following ground_truth format using fusion code between the annotation and OCR
• Save all the data in the following structure of dataset

python download_dataset.py --dataset_root_folder <output path to the dataset root folder> --tessdata_path <path to tessdata folder>

This repository assumes the following structure of dataset:

> tree dataset
dataset
├── test
│   ├── gts
│         ├── {image1}.json
│         ├── {image2}.json
│               .
│               .
│   ├── ocr
│         ├── {image1}.json
│         ├── {image2}.json
│               .
│               .
│   ├── images
│         ├── {image1}.png
│         ├── {image2}.png
│               .
│               .
│   ├── annotation
│         ├── {image1}.json
│         ├── {image2}.json
│               .
│               .
│                          
├── train
│   ├── gts
│         ├── {image1}.json
│         ├── {image2}.json
│               .
│               .
│   ├── ocr
│         ├── {image1}.json
│         ├── {image2}.json
│               .
│               .
│   ├── images
│         ├── {image1}.png
│         ├── {image2}.png
│               .
│               .
│   ├── annotation
│         ├── {image1}.json
│         ├── {image2}.json
│               .
│               .
│                          

• The structure of ground_truth json, which differentiate between 3 types of KVP: unkeyed, unvalued, kvp

{
  "kvps_list" : [
    {
      "type" : "unkeyed",
      "key" : {
        "text" : "{matching name from unkeyed type list}"
      },
      "value" : {
        "text" : "Associated Value Text 0",
        "bbox" : [x1 , y1 , x2 , y2]
      }
    },
    {
      "type" : "unvalued",
      "key" : {
        "text" : "Example key text 1",
        "bbox" : [x1 , y1 , x2 , y2]
      }
    },
    {
      "type" : "kvp",
      "key" : {
        "text" : "Example Key Text 2",
        "bbox" : [x1 , y1 , x2 , y2]
      },
      "value" : {
        "text" : "Associated Value Text 2",
        "bbox" : [x1 , y1 , x2 , y2]
      }
    }
  ]
}

3. View and Statistics utilities for the dataset, Train and Benchmark

3.1. Setting up the environment:

git clone git@github.com:IBM/KVP10k.git
cd KVP10k/

#for conda env                         | #for virtual env
conda create -n KVP10k python=3.11     | python -m venv KVP10k
conda activate KVP10k                  | source KVP10k/bin/activate

pip3 install --only-binary "matplotlib" -e '.'

3.2. View and Statistics utilities for the dataset

We provide in the repository utilities to view and get statistics on the dataset

1. Annotation Visualization

python visualize_annotation.py --image_file <path to image file> --annotation_file <path to annotation file>

2. groud_truth Visualization

python visualize_gt.py --image_file <path to image file> --gt_file <path to groud_truth file>

3. Dataset statistics

python dataset_statistics.py --dataset_root_path <path to dataset root folder> [--splits test train]

3.3. train

For training A100 with 80GB is needed

python train.py --save_path <Path for saving the checkpoint> --training_folder <Path for the training data>

3.3. Inference

For inference A100 with 80GB/40GB is needed

python inference.py  --pretrained <Path for savein the checkpoint>/last.ckpt --test_folder <Path for the test data> --save_path <Path for storing the inference jsons>

3.4. Benchmarking

python benchmark.py  --gt_folder <Path for the ground truth jsons> --inference_folder <Path for the inference jsons folder>