OARDoc

OARDoc is a Python toolkit for training, evaluating, exporting, and running document image rectification models through one API and command-line interface.

Trained model results

The figures and OARDoc rows show actual evaluation results and rectification outputs from models trained with OARDoc. Paper-reported rows reproduce the rounded values published for each model. Model weights are not included.

UVDoc

UVDoc validation examples: input image, predicted backward grid, ground-truth rectification, and predicted rectification.

Result	Epochs	Images	MS-SSIM ↑	AD ↓
OARDoc trained model	21	50	0.783579	0.119782
Paper-reported	21	50	0.784	0.122

DocTr++

DocTr++ validation examples for complete, partial, and out-of-frame document boundaries, including ground-truth and predicted grids and rectifications.

DocUNet

Result	Epochs	Images	OCR images	MSSIM ↑	LD ↓	ED ↓	CER ↓
OARDoc trained model	65	130	60	0.505877	7.609163	507.433333	0.191467
Paper-reported	65	130	60	0.51	7.52	447.47	0.1695

UDIR

Result	Epochs	Images	OCR images	MSSIM-M ↑	LD-M ↓	ED ↓	CER ↓
OARDoc trained model	65	195	70	0.488510	10.225263	1029.714286	0.346802
Paper-reported	65	195	70	0.45	12.47	666.49	0.2288

The arrows indicate metric direction. UDIR OCR values are raw diagnostics and are sensitive to unfilled output regions.

Installation

pip

Install the package and its core inference dependencies from a source checkout:

python -m pip install .

Add the training dependencies when preparing datasets, training models, or running metrics that need optional backends:

python -m pip install ".[train]"

Install export support when producing ONNX artifacts:

python -m pip install ".[export]"

uv

Create the project environment:

uv sync

Include training, export, and development tools when working on the repository:

uv sync --extra train --extra export --group dev

Run commands inside the managed environment with uv run, for example uv run pytest or uv run oardoc info model=uvdoc.yaml.

Usage

Python API

Load a checkpoint and rectify one image:

from oardoc import OARDoc

model = OARDoc("weights.pt")
results = list(model.predict(source="document.jpg", save=True))
print(results[0].save_dir)

Start a training run from a bundled model and dataset configuration:

from oardoc import OARDoc

model = OARDoc("uvdoc.yaml")
run = model.train(data="uvdoc.yaml", data_root="datasets")
print(run["last"])

Configuration names resolve from the bundled oardoc/configs/models/ and oardoc/configs/datasets/ directories. Explicit paths and configuration mappings are also accepted. Dataset paths in a bundled YAML are resolved relative to the data_root supplied by the caller.

Command-line interface

The CLI uses key=value arguments and parses values as YAML:

oardoc predict model=weights.pt source=document.jpg save=true
oardoc train model=uvdoc.yaml data=uvdoc.yaml data_root=datasets
oardoc val model=weights.pt data=uvdoc.yaml data_root=datasets
oardoc export model=weights.pt format=onnx

Run oardoc help for the workflow overview or oardoc help COMMAND for command-specific arguments.

Development

Run the test and static checks from the repository root:

uv run pytest
uv run ruff check .
uv run ruff format --check .

License

OARDoc is licensed under the Apache License 2.0. See LICENSE and THIRD_PARTY_NOTICES.md for details.