ebl-ai-api

Data+Code is part of Paper Sign Detection for Cuneiform Tablets from Yunus Cobanoglu, Luis Sáenz, Ilya Khait, Enrique Jiménez please contact us for access to data on Zenodoo and paper as it is under currently under review. See https://github.com/ElectronicBabylonianLiterature/cuneiform-ocr/blob/main/README.md for overview and general information of all repositories associated with the paper from above.

Ebl Ai Api

Server deploying a deep learning model for inference on detecting bounding boxes on cuneiform sign tablets For training please refer to cuneiform-sign-detection repo

Table of contents

• Setup
• Codestyle
• Running the tests
• Running the application
• Acknowledgements

Setup

Requirements:

• sudo apt-get install ffmpeg libsm6 libxext6  -y  
  (may be needed for open-cv python)

• Python 3.9

python3 -m venv ./.venv

pyre-configuration specifies paths specifically to .venv directory

pip3 install -r requirements

Run

python3 ebl_ai/check_installation.py

to check pytorch, mmcv, mmdet and mmocr installation.

Model

• Using FCENet (CVPR'2021)
• FCENet implementation: MMOCR
• FCENET with deconvolutions has slightly better performance.
• FCENET without deconvolutions.
• We use FCENET without deconvolutions and with Resnet-18 as Backbone (checkpoint and config specified in ./model directory)

Checkpoints and Model Config

• Checkpoint can be downloaded from https://drive.google.com/uc?id=17vISNEucGkSyZSfWfu5eOrBtut2wXRJ6 or from
• Model config is in https://github.com/ElectronicBabylonianLiterature/ebl-ai-api/blob/main/model/fcenet_no_dcvn.py

Running the tests

• Use command black ebl_ai_api to format code.
• Use command flake8 for linting.
• Use command pytest to run all tests.
• Use command pyre check for type-checking.

Running the server

waitress-serve --port=8001 --call ebl.app:get_app

Acknowledgements

• FCENET https://arxiv.org/abs/2104.10442
• Using https://github.com/open-mmlab/mmocr/blob/main/configs/textdet/fcenet/README.md (CVPR'2021)
• MMOCR https://github.com/open-mmlab/mmocr
• Deep learning of cuneiform sign detection with weak supervision using transliteration alignment https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0243039
• Synthetic Cuneiform Dataset (2000 Tablets) from https://github.com/cdli-gh/Cuneiform-OCR
• Annotated Tablets (75 Tablets) https://compvis.github.io/cuneiform-sign-detection-dataset/