# BaDLAD: A Large Multi-Domain Bengali Document Layout Analysis Dataset > BaDLAD is the first large-scale, human-annotated, multi-domain document layout analysis dataset for Bengali, covering 33,693 pages across six domains with 710K polygon annotations for four semantic unit types. - Authors: Bevaya AI Team - Published: 2023-03-01 - Updated: 2026-03-15 - Paper: Preprint (arXiv) - Paper: Published (ICDAR 2023) - Dataset: BaDLAD Dataset and Code - Source: Bevaya Labs, --- ## TL;DR BaDLAD is, to the authors' knowledge, the first large-scale, human-annotated, multi-domain document layout analysis dataset for Bengali (Bangla). It contains 33,693 document pages collected from six domains, with 710K polygon annotations across four semantic unit types: text-box, paragraph, image, and table. The authors benchmark several object detection architectures on the dataset to establish baseline performance. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$. The paper's primary contribution is the dataset itself: a new, large-scale, human-annotated corpus for Bengali document layout analysis (DLA). The curation pipeline, annotation protocol, and statistical characterization of the data constitute the bulk of the paper. **Secondary:** $\Psi\_{\text{Evaluation}}$. The paper benchmarks Faster R-CNN, Mask R-CNN, and YOLOv8 on the dataset, establishing baseline performance numbers and identifying where current English-DLA models fall short on Bengali documents. ## What is the motivation? Bengali (Bangla) is spoken by approximately 300 million people as a native language and another 37 million internationally, making it one of the most widely spoken languages in the world. Despite this, the infrastructure for processing Bengali documents lags considerably behind what is available for Latin-script languages. Bengali OCR has seen progress over the past decade, but document layout analysis (DLA) for Bengali remained severely underresourced at the time of this paper. The core problem is that DLA is a prerequisite for document transcription pipelines: before an OCR system can read a page, the page must be segmented into meaningful regions (paragraphs, text-boxes, images, tables). Existing large DLA datasets like PubLayNet and DocBank are either automatically annotated from born-digital PDFs or restricted to a single domain (scientific articles). Neither approach transfers well to Bengali documents, which are predominantly scanned or photographed originals and span a far more heterogeneous range of typographies, scripts, and layouts. Three compounding challenges make Bengali DLA particularly difficult: 1. Bengali is a non-Latin script with complex character composites, inflected forms, and conjunct consonants. Models trained on Latin-script data cannot be naively applied. 2. A large fraction of publicly available Bengali documents are scanned images of historical materials, some dating to the early 1800s. These exhibit degraded print quality, unusual typefaces, and inconsistent layouts. 3. Modern Bengali legal documents (property deeds) contain handwritten annotations, signatures, stamps, and free-form text interleaved with printed content, creating annotation challenges that rule-based systems handle poorly. Existing Bengali DLA resources were small and domain-specific, insufficient to train the data-hungry deep learning models that have become standard in the field. ## What is the novelty? The main contribution is BaDLAD itself: the first multi-domain, large-scale, human-annotated dataset for Bengali document layout analysis. Several aspects distinguish it from prior work: **Scale and human annotation.** BaDLAD contains 33,693 annotated document pages produced by a team of 13 trained annotators over four months, with three dedicated curators performing quality control and iteratively refining annotation guidelines. Annotation time was also recorded per sample, providing a proxy measure of layout complexity. **Polygon annotations.** Unlike axis-aligned bounding box annotations used in most DLA datasets, BaDLAD uses polygon annotations that follow the actual boundaries of text regions. This matters because Bengali documents frequently contain rotated text, irregular column arrangements, and text that does not align neatly to rectangular boxes. **Multi-domain coverage.** The dataset spans six distinct domains: - Books and magazines (largest component, 30,054 pages) - Government documents (1,285 pages) - Liberation war documents (1,004 pages) - Historical newspapers, pre-December 1971 (861 pages) - New newspapers (161 pages) - Property deeds (328 pages, held entirely in the test split to preserve confidentiality and to probe out-of-distribution performance) **Data-driven semantic unit discovery.** The four annotation categories were not chosen arbitrarily. The authors trained a self-supervised SwAV model on approximately 20,000 scraped Bengali PDFs and inspected the resulting cluster prototypes. The clusters revealed four major semantic categories, which became the annotation schema: text-box (isolated headings, captions, labels), paragraph (contiguous body text), image (any non-text visual element including logos and signatures), and table (structured row-column content, with internal elements also annotated according to their type). **Additional unannotated data.** The authors also release approximately 4 million unannotated document images to support unsupervised or semi-supervised DLA research. The dataset split is 60:40 train-to-test at the document level (not the page level), stratified by domain, ensuring that pages from the same source document land in the same split to prevent data leakage. ## What experiments were performed? The authors benchmark three standard object detection and segmentation architectures on BaDLAD: - **Faster R-CNN (F-RCNN)** with ResNet-50 backbone, trained with the Detectron framework, 10,000 iterations, learning rate 0.001 with decay 0.1, minibatch size 48. - **Mask R-CNN (M-RCNN)** with ResNet-50 backbone, same Detectron training configuration. A second variant (M-RCNN\*) uses ResNet-101, pre-trained on PubLayNet. - **YOLOv8** trained with the Ultralytics framework, 100 epochs, batch size 8, initial learning rate 0.01, weight decay 0.0005, warm-up over 3 iterations. Two variants are reported: object detection (bounding boxes) and segmentation (polygon masks). All models were pre-trained on one of: ImageNet, PubLayNet, or COCO. Performance is reported as mAP (COCO-style, IoU thresholds 0.50-0.95) computed per domain and per unit type (paragraph, text-box, image, table). The evaluation measures both bounding box prediction (for detection models) and segmentation mask quality (for Mask R-CNN and YOLOv8 segmentation). BaDLAD provides annotations in three formats: polygon, best-fit bounding box, and segmentation mask, allowing fair comparison across model types. ## What are the outcomes/conclusions? The benchmark results suggest that existing English DLA models, when fine-tuned on BaDLAD, learn meaningful Bengali layout features but leave substantial room for improvement: - YOLOv8 achieves the highest bounding box mAP at 70.46% averaged across domains and unit types. - In the segmentation setting, YOLOv8 achieves 35.69% mAP, while M-RCNN pre-trained on PubLayNet reaches 32.27%. - Models generally perform better on paragraphs and images than on text-boxes and tables. Tables are underrepresented in the training split (1,353 table annotations in training vs. a much larger number of paragraphs), which likely explains the weaker table detection. - Text-box detection is notably low given it is the second most frequent unit type. The authors attribute this to the structural diversity of text-box layouts (headers, captions, page numbers, etc.). - The property deeds domain, absent from training, is poorly handled by all models, demonstrating the expected out-of-distribution degradation. - Models show robustness in some challenging conditions: code-switched text (Bengali and English mixed), noisy scans, and partially torn pages. The authors acknowledge several limitations. The dataset is domain-imbalanced: books and magazines account for roughly 89% of pages. Several practically important domains (shopping receipts, ID cards, application forms) are absent. The dataset does not currently include word-level or character-level text content, which limits applicability to language-model-based DLA methods (such as LayoutLM-style approaches). Future work is expected to address this gap through word detection and recognition pipelines. ## Mapping to Matter vs. Meaning Framework How does BaDLAD's 4-class taxonomy map to our [Matter vs. Meaning framework](../matter-vs-meaning/)? | BaDLAD Class | Visual Primitive | Logical Role | Notes | | :--- | :--- | :--- | :--- | | **Paragraph** | `Text` | `Body` | Contiguous body text blocks. | | **Text-box** | `Text` | `Title` / `SectionHeader` / `Caption` / `PageHeader` / `PageFooter` / `PageNumber` | Isolated text fragments: headings, captions, labels, page numbers. BaDLAD does not distinguish between these roles; all non-paragraph text is collapsed here. | | **Image** | `Image` | `Figure` / `Logo` / `Stamp` | Any non-text visual element, including logos, signatures, and stamps. No sub-classification. | | **Table** | `Table` | *(primitive only)* | Structured row-column content. Internal elements are also annotated according to their type (text-box, image, etc.). | **Coverage gaps:** BaDLAD's 4-class schema is extremely coarse. It lacks explicit classes for `Formula`, `ListItem`, `Footnote`, `BibEntry`, `PageHeader`, `PageFooter`, and all form primitives (`Field`, `Selection`). The `Text-box` class conflates many distinct roles (headings, captions, page numbers) into a single category, requiring downstream disambiguation. The `Image` class similarly merges photographs, logos, and stamps. ## Reproducibility ### Models No new model architecture is introduced. The benchmarked models are standard implementations: - Faster R-CNN and Mask R-CNN via the Detectron framework (Facebook Research). - YOLOv8 via the Ultralytics library. - Backbone: ResNet-50 for most experiments; ResNet-101 for one M-RCNN variant. Model weights trained on BaDLAD are available via Google Drive, linked from the GitHub repository README (https://github.com/BengaliAI/badlad). The dataset is also hosted on Kaggle (https://www.kaggle.com/datasets/reasat/badlad-train). Note: the GitHub repository has no LICENSE file despite the paper claiming CC BY-SA 4.0. ### Algorithms - R-CNN models: 10,000 iterations, learning rate 0.001, decay 0.1, minibatch 48, warm-up 5 iterations. - YOLOv8: 100 epochs, batch size 8, initial learning rate 0.01, weight decay 0.0005, warm-up 3 iterations. - Pretrained weights from ImageNet, COCO, or PubLayNet depending on the variant. - Evaluation metric: COCO-style mAP at IoU 0.50-0.95. ### Data - 33,693 annotated pages; train split: 20,365 pages; test split: 13,328 pages. - Split is 60:40 stratified by domain (at document level, not page level). Property deeds are held entirely in the test set. - Polygon, bounding box, and segmentation mask formats are all provided. - An additional approximately 4 million unannotated images are released for unsupervised use. - License: CC BY-SA 4.0 (covering both the dataset and benchmark code, as stated in the paper). - The property deeds portion has been anonymized (PII removed) before inclusion. - Annotations were produced using the Labelbox platform by 13 trained annotators under iterative guidelines developed with three curators. ### Evaluation - Metric: mAP at IoU thresholds 0.50-0.95 (COCO protocol), reported per domain and per unit type. - No error bars or multi-run statistics are reported; results represent single training runs. - The test split includes the property deeds domain, which is absent from training, making it a partial out-of-distribution evaluation by design. - Benchmarks use models pre-trained on different source datasets (ImageNet, COCO, PubLayNet), so comparisons across pretraining regimes are not controlled for compute or data size. ### Hardware The paper does not report specific hardware configurations, GPU type or count, training wall-clock time, or memory requirements. Reproducibility of the benchmarks requires consulting the Detectron and Ultralytics documentation for typical resource estimates for ResNet-50/101 and YOLOv8 at the reported batch sizes. ## BibTeX ```bibtex @inproceedings{shihab2023badlad, title={BaDLAD: A Large Multi-Domain Bengali Document Layout Analysis Dataset}, author={Shihab, Md. Istiak Hossain and Hasan, Md. Rakibul and Emon, Mahfuzur Rahman and Hossen, Syed Mobassir and Ansary, Md. Nazmuddoha and Ahmed, Intesur and Rakib, Fazle Rabbi and Dhruvo, Shahriar Elahi and Dip, Souhardya Saha and Pavel, Akib Hasan and Meghla, Marsia Haque and Haque, Md. Rezwanul and Chowdhury, Sayma Sultana and Sadeque, Farig and Reasat, Tahsin and Humayun, Ahmed Imtiaz and Sushmit, Asif}, booktitle={Document Analysis and Recognition -- ICDAR 2023}, pages={326--340}, year={2023}, publisher={Springer}, doi={10.1007/978-3-031-41676-7\_19} } ```