# IndicDLP: A Foundational Dataset for Multi-Lingual and Multi-Domain Document Layout Parsing

> A 120K-page human-annotated dataset spanning 11 Indic languages, English, 12 document domains, and 42 region labels for document layout parsing.

- Authors: Bevaya AI Team
- Published: 2026-03-10
- Updated: 2026-03-12
- Paper: Preprint (arXiv) <https://arxiv.org/abs/2512.20236>
- Paper: Published (ICDAR 2025) <https://doi.org/10.1007/978-3-032-04614-7_2>
- Dataset: IndicDLP <https://huggingface.co/datasets/ai4bharat/indicdlp>
- Code: IndicDLP (GitHub) <https://github.com/AI4Bharat/IndicDLP>
- Model: YOLOv10x (fine-tuned) <https://github.com/AI4Bharat/IndicDLP>
- Model: DocLayout-YOLO (fine-tuned) <https://github.com/AI4Bharat/IndicDLP>
- Model: RoDLA (fine-tuned) <https://github.com/AI4Bharat/IndicDLP>
- Project: Project Page <https://indicdlp.github.io/>
- Source: Bevaya Labs, <https://labs.bevaya.ai/>

---


## TL;DR

IndicDLP is a 120K-page, human-annotated document layout dataset covering 11 Indic languages plus English, 12 document domains (newspapers, novels, textbooks, forms, etc.), and 42 region labels. Models trained on IndicDLP generalize better to other layout datasets than models trained on English-only resources. The paper received the ICDAR 2025 Best Student Paper Runner-Up award.

## What kind of paper is this?

**Dominant:** $\Psi\_{\text{Resource}}$: the headline contribution is the IndicDLP dataset itself: its scale, linguistic diversity, domain breadth, and annotation quality. The paper is organized around data curation, annotation workflow, and the resulting benchmark.

**Secondary:** $\Psi\_{\text{Evaluation}}$: a substantial portion of the paper benchmarks six detection models (YOLOv10, DiT, DocLayout-YOLO, DINO, RoDLA, Florence-2) on IndicDLP, evaluates cross-lingual zero-shot transfer, and measures the dataset's effectiveness as a pretraining source for other layout datasets.

## What is the motivation?

Existing large-scale layout datasets (PubLayNet, DocBank) offer volume but lack fine-grained labels and multilingual coverage. Manually annotated datasets (M6Doc, D4LA) provide richer labels and domain diversity but are too small to train robust models (9K and 12K pages, respectively). This gap is particularly acute for Indic documents, which encompass 11+ scripts with complex diacritics, conjunct characters, and historical typographic variation, yet have almost no representation in current layout benchmarks.

The authors also note a practical motivation: poor layout localization cascades into poor OCR for Indic scripts, where text recognition is already under-resourced. A high-quality, diverse layout dataset is therefore a prerequisite for downstream digitization pipelines.

## What is the novelty?

The core contribution is scale and diversity in a single human-annotated resource:

- **119,806 pages** with **1,856,241 annotated instances** (bounding boxes)
- **42 region labels** covering both physical (figure, table, paragraph) and logical (jumpline, flag, sidebar, hierarchical list levels) regions
- **11 Indic languages** (Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu) plus **English**
- **12 document domains**: newspapers, novels, magazines, manuals, textbooks, acts and rules, question papers, forms, brochures, notices, syllabi, research papers
- Documents span from pre-independence era to present day, including both digital-born and scanned sources

The annotation process involved a team of 50 individuals (3 to 4 annotators and 1 reviewer per language), guided by a 150-page annotation manual. Over 60% of validated documents were further checked by a team of 10 supercheckers, including the authors, to ensure cross-language and cross-domain consistency.

The annotation schema builds on M6Doc's guidelines but refines them: splitting generic labels (e.g., `list` into `ordered-list` and `unordered-list`; `caption` into `table-caption` and `figure-caption`), adding domain-specific labels (`jumpline`, `flag`, `placeholder-text`, `contact-info`, `website-link`), and introducing hierarchical nesting for section titles and list levels (up to three depth levels).

The authors also curate **UED-mini**, a merged and label-harmonized subset of DocLayNet and M6Doc (75K images, 25 unified labels), intended as a pretraining resource for Indic layout models.

## What experiments were performed?

### Baseline Benchmarking

All experiments use mean Average Precision as the primary metric, computed over IoU thresholds from 0.5 to 0.95 in steps of 0.05 (COCO-style):

$$\text{mAP@[.5:.95]} = \frac{1}{10} \sum\_{t \in \\{0.50, 0.55, \ldots, 0.95\\}} \text{AP}\_t$$

Six models were trained and evaluated on IndicDLP:

| Model | Params | mAP@50 | mAP@75 | mAP@[50:95] |
| :--- | :--- | :--- | :--- | :--- |
| **YOLOv10x** | 37M | 73.5 | **60.6** | **55.0** |
| DocLayout-YOLO | 20M | 73.5 | 60.0 | 54.5 |
| RoDLA | 342M | **74.1** | 57.7 | 53.1 |
| DINO | 46M | 69.7 | 53.4 | 49.2 |
| DiT | 86M | 67.2 | 51.8 | 47.8 |
| Florence-2 | 826M | 41.4 | 28.7 | 28.0 |

YOLOv10x offers the best efficiency-performance tradeoff. Of the six models, DiT, RoDLA, and DocLayout-YOLO were pretrained on document-specific data (IIT-CDIP, M6Doc, and DocSynth-300K, respectively), while the others used COCO or (in the case of Florence-2) FLD-5B natural images. Florence-2 underperforms despite its large capacity, likely because its pretraining emphasizes natural images over document layouts.

### Cross-Lingual Zero-Shot Transfer

Language-specific YOLOv10x models were trained on single-language subsets and evaluated on all other languages. Key findings:

- Zero-shot evaluation on unseen scripts drops 20-25 mAP points
- Languages sharing the same script (Hindi/Marathi via Devanagari; Assamese/Bengali via Bengali script) show modest cross-lingual gains, but still trail language-specific models
- A model trained on the full multilingual IndicDLP outperforms all single-language models, even those evaluated on their own language

### Domain-Wise Performance

Performance varies substantially by domain. Simpler layouts (novels, research papers) score highest. Complex multi-column layouts (newspapers, magazines) and domains with unique region types (forms, brochures) are harder.

### Pretraining Experiments

| Pretraining Source | Size | mAP@[50:95] |
| :--- | :--- | :--- |
| UED-mini | 75K | **57.7** (+1.9) |
| No pretraining (scratch) | - | 55.8 (baseline) |
| COCO | 118K | 55.0 (-0.8) |
| UED (full) | 785K | 53.9 (-1.9) |
| PubLayNet | 360K | 53.0 (-2.8) |

UED-mini (the curated, human-annotated subset) is the only pretraining source that improves over training from scratch. Larger but lower-quality or less diverse sources (PubLayNet, UED full) actually hurt performance. This suggests that annotation quality and domain diversity matter more than volume for pretraining.

### IndicDLP as a Pretraining Source

Conversely, pretraining on IndicDLP and fine-tuning on M6Doc, DocLayNet, and D4LA yields +2.8 mAP improvement on average, and faster convergence for DocLayNet. The dataset generalizes well beyond Indic layouts.

### Cross-Dataset Generalization

A YOLOv10x model trained on IndicDLP and evaluated on other datasets (M6Doc, D4LA, DocLayNet) shows the smallest cross-dataset performance drop of any training source tested, demonstrating that IndicDLP's diversity leads to better generalization on common physical region labels (paragraph, table, figure).

## What are the outcomes/conclusions?

1. **Scale and diversity in one package:** IndicDLP is, by a significant margin, the largest human-annotated layout dataset with multilingual and multi-domain coverage (120K pages vs. DocLayNet's 80K, with far broader language and domain representation).

2. **Script matters for layout:** Zero-shot cross-lingual results show a 20-25 mAP drop on unseen scripts, indicating that visual script properties do influence layout detection. This contradicts the common assumption that layout analysis is script-agnostic.

3. **Quality over quantity for pretraining:** The curated UED-mini (75K pages) outperforms the full UED (785K pages) as a pretraining source, reinforcing that careful curation and label harmonization matter more than raw scale.

4. **Strong generalization:** Models trained on IndicDLP transfer better to English-centric datasets than the reverse, suggesting the dataset's domain and layout diversity provides a robust visual prior.

5. **Limitations acknowledged by the authors:** The dataset uses axis-aligned bounding boxes only (no polygons or rotated boxes), which limits precision on skewed or rotated documents. The 42-class label set, while richer than most, still does not cover all possible regions as dedicated labels (e.g., QR codes and barcodes are subsumed under `figure`; poems are excluded entirely).

## Mapping to Matter vs. Meaning Framework

How does IndicDLP's 42-class taxonomy map to our [Matter vs. Meaning framework](../matter-vs-meaning/)? IndicDLP uses a hierarchical label set with depth-level variants for headings and lists.

### Headings & Structure

| IndicDLP Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **headline** | `Text` | `Title` | The main document title. |
| **sub-headline** | `Text` | `Subtitle` | Secondary headline. |
| **subsub-headline** | `Text` | `Subtitle` | Tertiary headline. |
| **chapter-title** | `Text` | `SectionHeader` | Chapter-level heading. |
| **section-title** | `Text` | `SectionHeader` | Section-level heading (H2 equivalent). |
| **sub-section-title** | `Text` | `SectionHeader` | Subsection heading (H3 equivalent). |
| **subsub-section-title** | `Text` | `SectionHeader` | Sub-subsection heading (H4 equivalent). |

### Content

| IndicDLP Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **paragraph** | `Text` | `Body` | Standard prose content. |
| **quote** | `Text` | `Blockquote` | Extended quotations. |
| **sidebar** | `Text` | `Sidebar` | Content outside the main text flow. |
| **figure** | `Image` | `Figure` | Visual content (charts, photos, diagrams not distinguished). |
| **table** | `Table` | *(primitive only)* | Tabular data. |
| **formula** | `Formula` | `DisplayEquation` | Mathematical notation. |
| **index** | `Table` | `Index` | Back-of-book index. |
| **table-of-contents** | `Table` | `TOC` | Table of contents. |

### Lists (Hierarchical)

| IndicDLP Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **ordered-list** | `Text` | `ListItem` | Numbered list. Container-level annotation. |
| **unordered-list** | `Text` | `ListItem` | Bulleted list. Container-level annotation. |
| **sub-ordered-list** | `Text` | `ListItem` | Nested numbered list (depth 2). |
| **sub-unordered-list** | `Text` | `ListItem` | Nested bulleted list (depth 2). |
| **subsub-ordered-list** | `Text` | `ListItem` | Nested numbered list (depth 3). |
| **subsub-unordered-list** | `Text` | `ListItem` | Nested bulleted list (depth 3). |

### Meta & Navigation

| IndicDLP Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **author** | `Text` | `Author` | Author name(s). |
| **dateline** | `Text` | `Dateline` | Date/location information. |
| **header** | `Text` | `PageHeader` | Running headers. |
| **footer** | `Text` | `PageFooter` | Running footers. |
| **folio** | `Text` | `PageNumber` | Page numbers. |
| **page-number** | `Text` | `PageNumber` | Explicit page index (distinct from folio in some contexts). |
| **footnote** | `Text` | `Footnote` | Bottom-of-page notes. |
| **figure-caption** | `Text` | `Caption` | Caption for figures. |
| **table-caption** | `Text` | `Caption` | Caption for tables. |
| **reference** | `Text` | `BibEntry` | Bibliography items. |
| **jumpline** | `Text` | `JumpLine` | "Continued on page X" navigation pointers. |
| **flag** | `Text` | `SectionHeader` | Newspaper section flags (e.g., "Sports", "Opinion"). |
| **advertisement** | `Image` | `Advertisement` | Commercial content. |
| **contact-info** | `Text` | `Address` | Physical or electronic address information. |
| **website-link** | `Text` | `Address` | URL or web reference. |
| **placeholder-text** | `Text` | `Instruction` | Template text awaiting replacement (forms context). |

### Exam/Question Paper Classes

| IndicDLP Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **first-level-question** | `Text` | `Body` | Top-level exam question. No direct M-vs-M equivalent; closest to M6Doc's "QuestionNumber" domain grammar. |
| **second-level-question** | `Text` | `Body` | Sub-question (e.g., part a, b, c). |
| **third-level-question** | `Text` | `Body` | Sub-sub-question. |
| **answer** | `Text` | `Body` | Answer text. |
| **options** | `Text` | `Option` | Multiple-choice options. Maps to the Forms `Option` role. |

**Coverage strengths:** IndicDLP is one of the richest taxonomies in the layout landscape. It provides hierarchical depth for both headings (3 levels) and lists (3 levels), separates figure/table captions, and includes domain-specific labels for newspapers (`flag`, `jumpline`) and exam papers (`first/second/third-level-question`, `answer`, `options`). The `JumpLine` and `Flag` roles are rarely covered by other datasets.

**Coverage gaps:** No explicit classes for `Code`, `OpticalCode` (barcodes/QR), form primitives (`Field`, `Selection`, `Checkbox`), `Logo`, `Stamp`, or `Watermark`. QR codes and barcodes are subsumed under `figure`. No relation annotations (reading order, hierarchy).

## Reproducibility

### Models

- Six architectures evaluated: YOLOv10x (37M), DocLayout-YOLO (20M), RoDLA (342M), DINO (46M), DiT (86M), Florence-2 (826M)
- Three fine-tuned checkpoints released (YOLOv10x, DocLayout-YOLO, RoDLA) via Zenodo, with inference code
- Original hyperparameters from each model's source paper were retained

### Algorithms

- All models trained using their default hyperparameters as specified in the original papers
- No additional training tricks (warmup, gradient clipping, mixed precision, data augmentation) are reported beyond what the original model papers specify
- Metric: mAP@[.5:.95] (COCO-style, thresholds from 0.5 to 0.95 in steps of 0.05)

### Data

- **Training split:** ~96K images (80%)
- **Validation split:** ~12K images (10%)
- **Test split:** ~12K images (10%)
- Stratified split preserving language and domain proportions
- Available on Hugging Face (gated access: requires agreeing to share contact info) and AIKosh (indiaai.gov.in)
- COCO-format JSON annotations (PNG images)
- Images scaled to max 1024px shortest side, 1333px longest side
- Annotation process: maker-checker workflow with 50 annotators/reviewers total (3-4 annotators + 1 reviewer per language); 10 supercheckers validated over 60% of documents; completed in approximately 8 months
- **License:** MIT (annotations and code); underlying document images are sourced from government bulletins, ebook stores, school textbooks, newspapers, arXiv, and Shodhganga, with per-source licenses not enumerated in the paper

### Evaluation

- Primary metric: mAP@[.5:.95]
- Cross-lingual zero-shot evaluation across all 12 languages
- Domain-wise mAP breakdown
- Cross-dataset transfer evaluation on M6Doc, DocLayNet, D4LA
- No error bars, significance tests, or multi-seed runs reported

### Hardware

- **Training:** 8x NVIDIA H100 GPUs
- Total GPU-hours not reported
- No inference latency or throughput benchmarks reported

## BibTeX

```bibtex
@inproceedings{nath2025indicdlp,
  title={IndicDLP: A Foundational Dataset for Multi-Lingual and Multi-Domain Document Layout Parsing},
  author={Nath, Oikantik and Kukkala, Sahithi and Khapra, Mitesh and Sarvadevabhatla, Ravi Kiran},
  booktitle={International Conference on Document Analysis and Recognition (ICDAR)},
  pages={23--39},
  year={2025},
  publisher={Springer},
  doi={10.1007/978-3-032-04614-7_2}
}
```