BaDLAD: A Large Multi-Domain Bengali Document Layout Analysis Dataset

ICDAR 2023 | Paper | Dataset | Competition | Project page

---

Abstract

While strides have been made in deep learning based Bengali Optical Character Recognition (OCR) in the past decade, the absence of large Document Layout Analysis (DLA) datasets has hindered the application of OCR in document transcription, e.g., transcribing historical documents and newspapers. Moreover, rule-based DLA systems that are currently being employed in practice are not robust to domain variations and out-of-distribution layouts. To this end, we present the first multidomain large Bengali Document Layout Analysis Dataset: BaDLAD. This dataset contains 33,695 human annotated document samples from six domains - i) books and magazines, ii) public domain govt. documents, iii) liberation war documents, iv) newspapers, v) historical newspapers, and vi) property deeds, with 710K polygon annotations for four unit types: text-box, paragraph, image, and table. Through preliminary experiments benchmarking the performance of existing state-of-the-art deep learning architectures for English DLA, we demonstrate the efficacy of our dataset in training deep learning based Bengali document digitization models.

---

Models

BaDLAD-trained checkpoints (Hugging Face):

Model	Hub	Notes
Mask R-CNN R50 (paper mask baseline)	bengaliAI/badlad-mrcnn-paper	Instance segmentation; leaderboard mask_map
Faster R-CNN R50 (paper bbox baseline)	bengaliAI/badlad-frcnn-paper	Bounding boxes only
YOLOv8m-seg	bengaliAI/badlad-yolov8m-seg	Layout seg; also used in bbocr

PubLayNet Detectron weights that appear beside older Drive dumps are LayoutParser base init checkpoints (not BaDLAD finetunes).

Leaderboard

Public mask-mAP board on the private BaDLAD paper hidden test. Test images and gold are not released. Scores are produced by Bengali.AI maintainers.

Board	https://huggingface.co/spaces/bengaliAI/badlad-leaderboard
Results table	https://huggingface.co/datasets/bengaliAI/badlad-results
Public train data	https://www.kaggle.com/datasets/reasat/badlad-train

How to get your model evaluated

1. Put a downloadable checkpoint on the Hub (or another stable URL).
2. Open a BaDLAD evaluation request issue.
3. Maintainers run decode on the closed test set and publish mask mAP only (predictions are not returned).

Maintainer tooling lives in leaderboard/.

uv sync                          # installs everything except detectron2
uv pip install 'git+https://github.com/facebookresearch/detectron2.git'  # needs CUDA toolkit

# publish a scored metrics.json (scores only — never hyps)
uv run python leaderboard/publish_results.py --metrics path/to/metrics.json --dry-run
uv run python leaderboard/publish_results.py --metrics path/to/metrics.json --publish

Full pinned versions in requirements.txt; uv.lock covers the uv-resolvable subset.

Mask R-CNN vs paper Table 3 (2026-07-25)

Checkpoint: bengaliAI/badlad-mrcnn-paper, score_thresh=0.05. Domain-wise mask AP (×100) compared to paper row M-RCNN | ImgNet | Mask:

Domain	n	P	Tx	I	Tb
Historical Newspapers	345	60.3 / 60.3	18.3 / 18.3	57.3 / 57.3	0.0 / 0.0
New Newspapers	65	41.4 / 41.4	13.1 / 13.2	45.2 / 45.2	1.9 / 1.9
Magazine and Books	11674	61.8 / 61.8	25.3 / 25.3	44.9 / 44.9	2.3 / 2.3
Liberation War Documents	402	71.1 / 71.2	26.8 / 26.8	1.1 / 1.0	40.1 / 40.1
Government Documents	514	49.4 / 39.1	23.7 / 18.7	26.1 / 19.4	5.1 / 3.7
Property Deeds	328	38.0 / 0.6	14.2 / 0.7	13.3 / 2.1	3.2 / 0.6