DocILE: Document Information Localization and Extraction Benchmark

Repository to work with the DocILE dataset and benchmark, used in the DocILE'23 CLEF Lab and ICDAR Competition. The competition deadline is on May 10 May 24, 2023 and comes with a $9000 prize pool.

The repository consists of:

• A python library, docile, making it easy to load the dataset, work with its annotations, pdfs and pre-computed OCR and run the evaluation.
• An interactive dataset browser notebook to visualize the document annotations, predictions and evaluation results.
• Training and inference scripts for Baseline methods provided together with pretrained checkpoints and results on test and validation sets.
• Tutorials to get quickly started with the repo.

Table of Contents:

• Download the dataset
• Installation
• Predictions format and running evaluation
• Tasks and evaluation metrics
• Pre-computed OCR
• Development instructions
• Dataset and benchmark paper and Supplementary Material

Download the dataset

First you need to obtain a secret token by following the instructions at https://docile.rossum.ai/. Then download and unzip the dataset by running:

./download_dataset.sh TOKEN annotated-trainval data/docile --unzip
./download_dataset.sh TOKEN test data/docile --unzip
./download_dataset.sh TOKEN synthetic data/docile --unzip
./download_dataset.sh TOKEN unlabeled data/docile --unzip

Run ./download_dataset.sh --help for more options, including how to only show urls (to download with a different tool than curl), how to download smaller unlabeled/synthetic chunks or unlabeled dataset without pdfs (with pre-computed OCR only).

You can also work with the zipped datasets when you turn off image caching (check Load and sample dataset tutorial for details).

Installation

Option 1: Install as a library

Install the library with:

pip install docile-benchmark

To convert pdfs into images, the library uses https://github.com/Belval/pdf2image. On linux you might need to install:

apt install poppler-utils

And on macOS:

brew install poppler

Now you have all dependencies to work with the dataset annotations, pdfs, pre-comptued OCR and to run the evaluation. You can install extra dependencies by running the following (although using one of the provided dockerfiles, as explained below, might be easier in this case):

pip install "docile-benchmark[interactive]"
pip install "docile-benchmark[ocr]"

The first line installs additional dependencies allowing you to use the interactive dataset browser in docile/tools/dataset_browser.py and the tutorials. The second line let's you rerun the OCR predictions from scratch (e.g., if you'd like to run it with different parameters) but to make it work, you might need additional dependencies on your system. Check https://github.com/mindee/doctr for the installation instructions (for pytorch).

Option 2: Use docker

There are two Dockerfiles available with the dependencies preinstalled:

• Dockerfile is a lighter, CPU-only, version with all necessary dependencies to use the dataset with the pre-computed OCR and interactive browser.
• Dockerfile.gpu has CUDA GPU support and contains additional dependencies needed to recompute OCR predictions from scratch (not needed for standard usage).

You can use docker compose to manage the docker images. First update the settings in docker-compose.yml and the port for jupyter lab in .env. Then build the image with:

docker compose build jupyter[-gpu]

where jupyter uses Dockerfile and jupyter-gpu uses Dockerfile.gpu. You can then start the jupyter server:

docker compose up -d jupyter[-gpu]

Jupyter lab can be then accessed at https://127.0.0.1:${JUPYTER_PORT} (retrieve the token from logs with docker compose logs jupyter[-gpu]). You can also login to the container with:

docker compose exec jupyter bash

After that run poetry shell to activate the virtual environment with the docile library and its dependencies installed.

Predictions format and running evaluation

To evaluate predictions for tasks KILE or LIR on the test set, you have to make a submission to the benchmark on the Robust Reading Competition website. To evaluate on the validation set, use the following command:

docile_evaluate \
  --task LIR \
  --dataset-path path/to/dataset[.zip] \
  --split val \
  --predictions path/to/predictions.json \
  --evaluate-x-shot-subsets "0,1-3,4+" \  # default, show evaluation for 0-shot, few-shot and many-shot layout clusters
  --evaluate-synthetic-subsets \  # optional, show evaluation on layout clusters with available synthetic data
  --evaluate-fieldtypes \  # optional, show breakdown per fieldtype
  --evaluate-also-text \  # optional, evaluate if the text prediction is correct
  --store-evaluation-result LIR_val_eval.json  # optional, it can be loaded in the dataset browser

Run docile_evaluate --help for more information on the options. You can also run docile_print_evaluation_report --evaluation-result-path LIR_val_eval.json to print the results of a previously computed evaluation. It is also possible to run evaluation or load the evaluation result directly from code which provides even more options, such as computing metrics just for a single document, specific subset of documents (i.e., documents belonging to the same layout cluster) etc. Check docile/evaluation/evaluate.py module for more details.

Predictions need to be stored in a single json file (for each task separately) containing a mapping from docid to the predictions for that document, i.e.:

{
    "docid1": [
        {
            "page": 0,
            "bbox": [0.2, 0.1, 0.4, 0.5],
            "fieldtype": "line_item_order_id",
            "line_item_id": 3,
            "score": 0.8,
            "text": "Order 38",
            "use_only_for_ap": true
        },
        "..."
    ],
    "docid2": [{"...": "..."}, "..."],
    "..."
}

Explanation of the individual fields of the predictions:

• page: page index (from zero) the prediction belongs to
• bbox: relative coordinates (from 0 to 1) representing the left, top, right, bottom sides of the bbox, respectively
• fieldtype: the fieldtype (sometimes called category or key) of the prediction
• line_item_id: ID of the line item. This should be a different number for each line item, the order does not matter. Omit for KILE predictions.
• score [optional]: the confidence for this prediction, can be omitted (in that case predictions are taken in the order in which they are stored in the list)
• text [optional]: text of the prediction, evaluated in a secondary metric only (when --evaluate-also-text is used)
• use_only_for_ap [optional, default is False]: only use the prediction for AP metric computation, not for f1, precision and recall (useful for less confident predictions).

You can use Field class to work with the predictions in code, in which case you can store them in a json file in the required format by using the docile.dataset.store_predictions function. You can also use the BBox class to easily convert between relative and absolute coordinates (relative coordinates are used everywhere by default).

Use docile/tools/print_results.py to display results of your predictions formatted in a table (choosing from table styles like github, latex or csv).

Tasks and evaluation metrics

The DocILE benchmark comes with two tracks, Key Information Localization and Extraction (KILE) and Line Item Recognition (LIR), as described in the dataset and benchmark paper. In both tracks, the goal is to localize key information of pre-defined categories, i.e., for each document, detect fields with their fieldtype, location (page and bbox) and optionally text. For LIR, fields have to be additionally grouped into line items. A Line Item (LI) is a tuple of fields (e.g., description, quantity, and price) describing a single object instance to be extracted, e.g., a row in a table.

Evaluated metrics are Average Precision (AP), F1, precision and recall, computed over all fields across all documents and field types (micro-average). A predicted and a ground truth field can be matched if they have the same fieldtype and if they are in the same location, which is decided by the overlap of their bounding boxes with the provided OCR words, described more precisely in the dataset and benchmark paper:

For LIR, the fields have to additionally belong to corresponding Line Items. The mapping between line_item_id for predicted and ground truth fields is decided by taking the maximum matching which maximizes the total number of matched fields, when considering only fields with use_only_for_ap=False (because use_only_for_ap=True can be used to include also less confident predictions for AP computation which could negatively affect this matching).

For both KILE and LIR, the metrics are then computed like this:

• For each document page, find matching between predicted and ground truth fields. If more predicted fields match the same ground truth field, the one with higher score is used (but primarily the one with use_only_for_ap=False).
• Order all predictions by descending score (but again, predictions with use_only_for_ap=True come last). Break ties by the original order in which predictions were provided. See evaluate.py for the precise ordering rules. Notice that the order does not influence F1, precision and recall but it influences AP.
• Compute F1, precision and recall by counting number of true positives and false positives/negatives and AP by iteratively adding predictions and updating precision and recall. We use the COCO style of AP where "gaps are filled", i.e., the precision-recall curve becomes non-increasing.

Pre-computed OCR

Pre-computed OCR is provided with the dataset. The prediction was done using the DocTR library. On top of that, word boxes were snapped to text (check the code in docile/dataset/document_ocr.py). These snapped word boxes are used in evaluation.

To get the OCR words for a single page, call document.ocr.get_all_words(page), optionally passing snapped=True to get the snapped bounding boxes. In some rare cases, the list of words might be empty (when the page is blank). Also notice the predicted words are grouped into blocks. For some models it might be beneficial to re-order the words primarily by lines.

While it should not be needed, it is possible to (re)generate OCR from scratch (including the snapping) with the provided Dockerfile.gpu. Just delete DATASET_PATH/ocr directory and then access the ocr for each document and page with document.ocr.get_all_words(page, snapped=True).

Development instructions

For development, install poetry and run poetry install. Start a shell with the virtual environment activated with poetry shell. No other dependencies are needed to run pre-commit and the tests. It's recommended to use docker (as explained above) if you need the extra (interactive or ocr) dependencies.

Install pre-commit with pre-commit install (don't forget you need to prepend all commands with poetry run ... if you did not run poetry shell first).

Run tests by calling pytest tests.

Dataset and benchmark paper and Supplementary Material

The dataset, the benchmark tasks and the evaluation criteria are described in detail in the dataset paper which was accepted to ICDAR 2023. The provided link is to arXiv version that includes Supplementary Material. To cite the dataset, please use the following BibTeX entry:

@misc{simsa2023docile,
    title={{DocILE} Benchmark for Document Information Localization and Extraction},
    author={{\v{S}}imsa, {\v{S}}t{\v{e}}p{\'a}n and {\v{S}}ulc, Milan and U{\v{r}}i{\v{c}}{\'a}{\v{r}}, Michal and Patel, Yash and Hamdi, Ahmed and Koci{\'a}n, Mat{\v{e}}j and Skalick{\`y}, Maty{\'a}{\v{s}} and Matas, Ji{\v{r}}{\'\i} and Doucet, Antoine and Coustaty, Micka{\"e}l and Karatzas, Dimosthenis},
    url = {https://arxiv.org/abs/2302.05658},
    journal={arXiv preprint arXiv:2302.05658},
    year={2023}
}