# M6Doc: Multi-Format, Multi-Type, Multi-Layout, Multi-Language, Multi-Annotation Dataset

> Analysis of M6Doc, a fine-grained layout analysis dataset with 74 classes across 9,080 pages. We examine its annotation methodology, class distribution, and the TransDLANet baseline proposed alongside it.

- Authors: Bevaya AI Team
- Published: 2023-06-01
- Updated: 2026-03-15
- Paper: Published (CVPR 2023) <https://doi.org/10.1109/CVPR52729.2023.01453>
- Paper: Preprint (arXiv) <https://arxiv.org/abs/2305.08719>
- Code: HCIILAB/M6Doc <https://github.com/HCIILAB/M6Doc>
- Dataset: Test Set (OneDrive) <https://1drv.ms/u/s!AuZg70Qbn4fEiVGjzKiExS5P65rV?e=EOIxKe>
- Dataset: Train/Val Set (request via GitHub) <https://github.com/HCIILAB/M6Doc>
- Source: Bevaya Labs, <https://labs.bevaya.ai/>

---


## TL;DR

M6Doc is a manually annotated document layout dataset covering 9,080 pages across seven document types, three formats (PDF, scanned, photographed), and two languages, with 74 fine-grained annotation categories and 237,116 instances. The paper also introduces TransDLANet, a transformer-based instance segmentation model, which achieves 64.5% mAP on the dataset.

## What kind of paper is this?

**Dominant:** $\Psi\_{\text{Resource}}$: the primary contribution is the M6Doc dataset itself, including its collection pipeline, annotation schema, and 74-class label taxonomy.

**Secondary:** $\Psi\_{\text{Method}}$: TransDLANet is a purpose-built baseline model introduced alongside the dataset, representing a real methodological contribution rather than just an evaluation vehicle.

**Secondary:** $\Psi\_{\text{Evaluation}}$: the paper benchmarks eleven object detection and five instance segmentation baselines against the dataset, positioning M6Doc as a community benchmark.

## What is the motivation?

The field's trajectory runs from PRImA (2009, 305 scanned pages of magazines and scientific articles, widely considered the first real-world DLA dataset) through the large automatically annotated PDF corpora of the late 2010s: PubLayNet (360k pages, 5 classes), DocBank (500k pages, 13 classes), and DocLayNet (80k pages, 11 classes). M6Doc positions itself against several structural limitations shared by that generation:

1. **Format monoculture.** All large public datasets consist entirely of PDF documents. Models trained on them tend to fail on scanned or photographed pages, which dominate real-world document workflows.
2. **Type monoculture.** Most datasets are drawn exclusively from scientific articles, which use highly regularized two-column or single-column templates. Magazines, newspapers, textbooks, and exam papers have fundamentally different visual grammars.
3. **Coarse taxonomies.** Five-class taxonomies (Text, Title, Table, Figure, List) are sufficient for document retrieval but insufficient for reconstruction, accessibility pipelines, or editorial understanding. Distinguishing a `kicker` from a `headline` or a `footnote` from a `marginal note` requires categories that these datasets lack.
4. **Language coverage.** Nearly all large-scale public data is English. Non-English work exists (BCE-Arabic reaches 9k pages of scanned Arabic books) but remains narrow in scope and format. Document layout varies with writing direction, script density, and typographic conventions.

### The "M6" Properties

The dataset name encodes six design properties:

| # | Property | Details |
| :--- | :--- | :--- |
| 1 | Multi-Format | Scanned, photographed, and born-digital PDF |
| 2 | Multi-Type | Scientific articles, textbooks, books, test papers, magazines, newspapers, notes |
| 3 | Multi-Layout | Rectangular, Manhattan, non-Manhattan, multi-column Manhattan |
| 4 | Multi-Language | Chinese and English |
| 5 | Multi-Annotation Category | 74 annotation classes |
| 6 | Modern Documents | Contemporary layouts drawn from current publications; explicitly excludes historical, archival, or pre-digital documents (e.g., no early-print OCR corpora) |

## What is the novelty?

### The Dataset

M6Doc contains 9,080 manually annotated pages with 237,116 total instances, split 6:1:3 across training (5,448 pages, 143,040 annotations), validation (908 pages, 23,210 annotations), and test (2,724 pages, 70,866 annotations).

**Document type composition:** scientific article (11%), textbook (23%), test paper (22%), magazine (22%), newspaper (11%), note (5.5%), book (5.5%).

**Format composition:** PDF (64%), scanned (31%), photographed (5%).

**Sources:** arXiv (scientific articles), the Chinese People's Daily website (newspapers), scanned textbooks across nine school subjects, handwritten student notes, and photographed books. Magazines were collected as PDFs from a mix of Chinese and English publishers (including The New Yorker, Scientific American, and Global Science); the paper also cites VKontakte as a general source, though which specific content came from there is not entirely clear. As with any dataset that aggregates third-party publications, the copyright ownership of individual items may not rest with the dataset creators, which creates some uncertainty about the scope of the CC-BY-NC-ND-4.0 license.

The 74-class taxonomy is, to our knowledge, the most granular publicly released layout annotation schema. It covers editorial structures specific to print journalism (kicker, dateline, jump line, byline, mugshot), academic structures (formula, algorithm, reference, footnote, endnote), exam-specific structures (question numbers at three levels, answer, option, matching, blank), and general structures (section titles at four levels, ordered/unordered lists, figures, tables, and navigation elements). Compared to DocLayNet's 11 classes, this is roughly a $7\times$ increase in granularity.

A key design note: list items are grouped into container blocks (ordered list, unordered list) rather than annotated individually, which differs from DocLayNet's per-line list-item convention.

### Full Class Distribution

| Category | Train | Train % | Val | Val % | Test | Test % |
|--------|--------:|--------:|------:|------:|-------:|-------:|
| _background_ | 0 | 0.000 | 0 | 0.000 | 0 | 0.000 |
| QR code | 59 | 0.041 | 15 | 0.065 | 23 | 0.032 |
| advertisement | 257 | 0.180 | 45 | 0.194 | 145 | 0.205 |
| algorithm | 12 | 0.008 | 3 | 0.013 | 12 | 0.017 |
| answer | 165 | 0.115 | 30 | 0.129 | 77 | 0.109 |
| author | 2,424 | 1.695 | 403 | 1.736 | 1,188 | 1.676 |
| barcode | 10 | 0.007 | 1 | 0.004 | 3 | 0.004 |
| bill | 3 | 0.002 | 2 | 0.009 | 3 | 0.004 |
| blank | 189 | 0.132 | 58 | 0.250 | 90 | 0.127 |
| bracket | 863 | 0.603 | 164 | 0.707 | 273 | 0.385 |
| breakout | 411 | 0.287 | 72 | 0.310 | 188 | 0.265 |
| byline | 1,276 | 0.892 | 185 | 0.797 | 660 | 0.931 |
| caption | 3,508 | 2.452 | 605 | 2.607 | 1,766 | 2.492 |
| catalogue | 39 | 0.027 | 10 | 0.043 | 19 | 0.027 |
| chapter title | 245 | 0.171 | 33 | 0.142 | 124 | 0.175 |
| code | 62 | 0.043 | 7 | 0.030 | 31 | 0.044 |
| correction | 9 | 0.006 | 1 | 0.004 | 6 | 0.008 |
| credit | 1,523 | 1.065 | 255 | 1.099 | 728 | 1.027 |
| dateline | 901 | 0.630 | 140 | 0.603 | 482 | 0.680 |
| drop cap | 414 | 0.289 | 71 | 0.306 | 234 | 0.330 |
| editor's note | 39 | 0.027 | 4 | 0.017 | 9 | 0.013 |
| endnote | 35 | 0.024 | 4 | 0.017 | 19 | 0.027 |
| examinee information | 8 | 0.006 | 2 | 0.009 | 6 | 0.008 |
| fifth-level title | 13 | 0.009 | 2 | 0.009 | 20 | 0.028 |
| figure | 7,614 | 5.323 | 1,242 | 5.351 | 3,762 | 5.309 |
| first-level question number | 5,669 | 3.963 | 930 | 4.007 | 2,740 | 3.866 |
| first-level title | 586 | 0.410 | 81 | 0.349 | 292 | 0.412 |
| flag | 30 | 0.021 | 5 | 0.022 | 12 | 0.017 |
| folio | 1,442 | 1.008 | 213 | 0.918 | 685 | 0.967 |
| footer | 1,984 | 1.387 | 310 | 1.336 | 987 | 1.393 |
| footnote | 295 | 0.206 | 49 | 0.211 | 139 | 0.196 |
| formula | 13,090 | 9.151 | 2,058 | 8.867 | 6,191 | 8.736 |
| fourth-level section title | 15 | 0.010 | 3 | 0.013 | 19 | 0.027 |
| fourth-level title | 70 | 0.049 | 13 | 0.056 | 66 | 0.093 |
| header | 1,877 | 1.312 | 297 | 1.280 | 969 | 1.367 |
| headline | 4,115 | 2.877 | 643 | 2.770 | 1,981 | 2.795 |
| index | 214 | 0.150 | 36 | 0.155 | 100 | 0.141 |
| inside | 16 | 0.011 | 1 | 0.004 | 5 | 0.007 |
| institute | 60 | 0.042 | 9 | 0.039 | 28 | 0.040 |
| jump line | 381 | 0.266 | 63 | 0.271 | 180 | 0.254 |
| kicker | 516 | 0.361 | 91 | 0.392 | 257 | 0.363 |
| lead | 664 | 0.464 | 109 | 0.470 | 285 | 0.402 |
| marginal note | 238 | 0.166 | 37 | 0.159 | 101 | 0.143 |
| matching | 7 | 0.005 | 1 | 0.004 | 8 | 0.011 |
| mugshot | 73 | 0.051 | 11 | 0.047 | 46 | 0.065 |
| option | 3,198 | 2.236 | 515 | 2.219 | 1,577 | 2.225 |
| ordered list | 1,012 | 0.707 | 172 | 0.741 | 510 | 0.720 |
| other question number | 42 | 0.029 | 3 | 0.013 | 31 | 0.044 |
| page number | 4,782 | 3.343 | 803 | 3.460 | 2,383 | 3.363 |
| paragraph | 65,642 | 45.891 | 10,575 | 45.562 | 33,069 | 46.664 |
| part | 524 | 0.366 | 89 | 0.383 | 283 | 0.399 |
| play | 10 | 0.007 | 3 | 0.013 | 2 | 0.003 |
| poem | 98 | 0.069 | 18 | 0.078 | 33 | 0.047 |
| reference | 149 | 0.104 | 23 | 0.099 | 62 | 0.087 |
| sealing line | 3 | 0.002 | 2 | 0.009 | 5 | 0.007 |
| second-level question number | 2,773 | 1.939 | 377 | 1.624 | 1,330 | 1.877 |
| second-level title | 273 | 0.191 | 48 | 0.207 | 140 | 0.198 |
| section | 2,508 | 1.753 | 408 | 1.758 | 1,228 | 1.733 |
| section title | 897 | 0.627 | 171 | 0.737 | 442 | 0.624 |
| sidebar | 54 | 0.038 | 10 | 0.043 | 27 | 0.038 |
| sub section title | 567 | 0.396 | 107 | 0.461 | 269 | 0.380 |
| subhead | 1,998 | 1.397 | 394 | 1.698 | 1,069 | 1.508 |
| subsub section title | 101 | 0.071 | 21 | 0.090 | 71 | 0.100 |
| supplementary note | 986 | 0.689 | 158 | 0.681 | 487 | 0.687 |
| table | 821 | 0.574 | 146 | 0.629 | 409 | 0.577 |
| table caption | 287 | 0.201 | 41 | 0.177 | 143 | 0.202 |
| table note | 8 | 0.006 | 2 | 0.009 | 5 | 0.007 |
| teasers | 32 | 0.022 | 7 | 0.030 | 7 | 0.010 |
| third-level question number | 240 | 0.168 | 36 | 0.155 | 102 | 0.144 |
| third-level title | 146 | 0.102 | 44 | 0.190 | 94 | 0.133 |
| title | 201 | 0.141 | 35 | 0.151 | 100 | 0.141 |
| translator | 73 | 0.051 | 11 | 0.047 | 38 | 0.054 |
| underscore | 3,687 | 2.578 | 590 | 2.542 | 1,717 | 2.423 |
| unordered list | 497 | 0.347 | 84 | 0.362 | 271 | 0.382 |
| weather forecast | 10 | 0.007 | 3 | 0.013 | 3 | 0.004 |
| **Total** | **143,040** | **100** | **23,210** | **100** | **70,866** | **100** |

### TransDLANet

Alongside the dataset, the authors propose TransDLANet, a transformer-based instance segmentation model. It is derived from the ISTR framework with three modifications: (1) a Transformer encoder without positional encoding for self-attention over query embeddings; (2) an adaptive element matching mechanism to improve recall by better aligning queries to ground-truth instances; and (3) three shared-parameter MLP branches for joint classification, bounding box regression, and segmentation mask prediction. The model iteratively refines its output over $K$ rounds.

## What experiments were performed?

### Annotation Methodology

47 annotators followed a purpose-written guideline of over 170 pages. Labels were defined using a "commonality vs. specificity" principle: labels were unified across document types where possible, and type-specific labels were added only where the distinction carried semantic value. Annotation files use MS COCO format. Final quality control was performed by the authors.

The authors report several sources of inter-annotator ambiguity in the guideline: determining table vs. paragraph boundaries when borders are absent; deciding whether a figure containing sub-figures should be annotated as one region or several; and distinguishing list items from paragraphs when no explicit marker is present. Resolving these consistently required explicit annotation rules.

### Baseline Benchmarks on M6Doc

The paper evaluates eleven object detection methods and five instance segmentation methods. The paper reports both object detection mAP and instance segmentation mAP separately. Selected results on the M6Doc test set (COCO mAP 0.50:0.95):

| Method | Backbone | Det. mAP | Seg. mAP |
| :--- | :--- | ---: | ---: |
| RetinaNet | ResNet-101 | 21.4 | 21.0 |
| YOLOv3 | DarkNet-53 | 59.8 | -- |
| FCOS | ResNet-101 | 40.6 | 39.3 |
| Faster R-CNN | ResNet-101 | 49.0 | 47.8 |
| Cascade R-CNN | ResNet-101 | 54.1 | 52.7 |
| Mask R-CNN | ResNet-101 | 40.1 | 39.7 |
| Cascade Mask R-CNN | ResNet-101 | 54.4 | 52.9 |
| HTC | ResNet-101 | 58.2 | 57.1 |
| SOLOv2 | ResNet-101 | 46.8 | 48.3 |
| Deformable DETR | ResNet-101 | 57.2 | 55.6 |
| **TransDLANet (proposed)** | ResNet-101 | **64.5** | **63.4** |

Note that for detection-only methods (e.g., Faster R-CNN), the segmentation mAP is computed using the bounding box as a proxy mask. For segmentation-only methods (e.g., SOLOv2), the detection mAP uses the minimum bounding rectangle of the predicted mask. All anchor-based methods use a non-default anchor ratio set of [0.0625, 0.125, 0.25, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0] to accommodate M6Doc's wide aspect ratio diversity.

A notable result: Mask R-CNN (40.1 detection mAP) substantially underperforms Faster R-CNN (49.0 detection mAP), despite Mask R-CNN typically performing comparably or better on standard benchmarks. The authors note that pixel-based segmentation degrades performance when documents have complex layouts, a finding they also observe on DocLayNet. Meanwhile, SOLO reaches only 38.7 segmentation mAP. TransDLANet builds on ISTR, and the transformer encoder and adaptive matching exist specifically to address ISTR's fixed-query collapse when multiple queries converge on the same dense region.

The authors also evaluate TransDLANet on DocLayNet and PubLayNet. On DocLayNet, TransDLANet achieves 72.3% mAP, below the YOLOv5x6 baseline at 76.8%. On PubLayNet, it achieves 94.5% mAP, comparable to but slightly below VSR at 95.7%. The authors attribute the DocLayNet gap to a fixed-query design that misses instances when multiple queries collapse onto a single region.

### Cross-Dataset Transfer Experiments

To demonstrate the value of format diversity, the authors trained models on DocBank, PubLayNet, and DocLayNet and evaluated on M6Doc's scanned and photographed subsets. Models trained solely on PDF data consistently fail to detect instances in scanned handwritten notes and photographed books, reflecting domain gap from background variation, tilt, and brightness shifts. Models trained on M6Doc generalize better to those formats. Conversely, models trained on M6Doc do not transfer well to DocLayNet (and vice versa), which the authors attribute to the two datasets drawing from entirely different source corpora.

## What are the outcomes/conclusions?

M6Doc demonstrates that increasing label granularity, even at the cost of raw dataset size, can improve a model's ability to extract fine-grained logical structure. A Faster R-CNN trained on M6Doc's 600-page scientific article subset produces more accurate results on formula and section-level distinctions than the same architecture trained on DocBank, which has 500k pages but only 13 coarse classes.

The dataset is genuinely hard: the strongest baseline excluding TransDLANet (HTC) reaches only 58.2% detection mAP. Failure modes the authors identify include: dense or skewed instances that overlap significantly, making anchor-based suppression lossy; large-aspect-ratio regions like newspaper advertisements that fall outside standard anchor configurations; and handwritten notes, which remain the most difficult subset for all evaluated models.

**Critical observations:**

- The "Large-Scale" framing in the title is hard to justify. At 9,080 pages, M6Doc is substantially smaller than DocLayNet (80k), PubLayNet (360k), or DocBank (500k). Its value is in annotation granularity and format diversity, not volume.
- The class imbalance is extreme. `paragraph` accounts for roughly 46% of all instances while several classes have single-digit counts in validation (e.g., `matching`: 1 val instance; `barcode`: 1 val instance; `bill`: 2 val instances). Standard COCO mAP weights all classes equally, so headline numbers are dominated by the high-frequency classes and tell you almost nothing about model performance on the rare but semantically distinctive ones.
- The CC-BY-NC-ND-4.0 license prevents commercial use and prohibits derivative datasets, which limits applicability for production-oriented work.
- The Train/Val split is not publicly downloadable without requesting access through the repository. Only the test set is on open download.
- No training or inference code is provided. The baselines must be reproduced independently, with the anchor ratio adjustment ([0.0625, 0.125, ..., 16.0]) being a non-obvious configuration change needed for anchor-based methods to handle M6Doc's wide aspect ratio diversity.

## Mapping to Unified Taxonomy

M6Doc's granularity stress-tests any unified taxonomy. The tables below map each of its 74 classes to our [Matter/Meaning annotation standard](../matter-vs-meaning), noting where M6Doc distinguishes concepts that are usually merged. (Note: that page is a work in progress; check current availability before linking externally.)

### Hierarchy and Text Flow

| M6Doc Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **title** | `Text` | `Title` | Document-level title. |
| **first-level title** through **fifth-level title** | `Text` | `Title` / `SectionHeader` | M6Doc splits academic headers by level. |
| **section title** | `Text` | `SectionHeader` | Generic section divider. |
| **sub section title** | `Text` | `SectionHeader` | Second-level academic section. |
| **subsub section title** | `Text` | `SectionHeader` | Third-level academic section. |
| **fourth-level section title** | `Text` | `SectionHeader` | Fourth-level academic section. |
| **second-level title** / **third-level title** | `Text` | `SectionHeader` | Parallel hierarchy for non-article documents. |
| **chapter title** | `Text` | `SectionHeader` | Book-level chapter division. |
| **section** / **part** | `Text` | `SectionHeader` | High-level structural divisions (books, textbooks). |
| **headline** | `Text` | `Title` | Newspaper article title. |
| **subhead** | `Text` | `SectionHeader` | Secondary headline (newspaper). |
| **kicker** | `Text` | `SectionHeader` | Introductory label above the headline. |
| **paragraph** | `Text` | `Body` | Standard prose body. |
| **drop cap** | `Text` | `Body` | Typographic first letter; semantically part of the paragraph. |
| **breakout** | `Text` | `Blockquote` | Pull quote or highlighted excerpt. |
| **lead** | `Text` | `Abstract` | Introductory summary paragraph in news articles. |
| **supplementary note** | `Text` | `Body` | Supplementary explanatory text. |
| **jump line** | `Text` | `JumpLine` | "Continued on page X" cross-reference. |
| **byline** / **author** | `Text` | `Author` | Attribution lines. |
| **institute** / **translator** | `Text` | `Author` | Secondary attribution blocks. |

### Specialized Content

| M6Doc Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **advertisement** | `Image` | `Advertisement` | Distinct commercial region. |
| **code** | `Text` | `Code` | Source code blocks. |
| **algorithm** | `Text` | `Code` | Pseudocode or algorithmic descriptions. |
| **poem** | `Text` | `Body` | Verse with significant whitespace. |
| **play** | `Text` | `Body` | Dialogue-heavy script text. |
| **editor's note** | `Text` | `Abstract` | Editorial comment or contextual summary. |
| **correction** | `Text` | `Body` | Published correction notice. |
| **bill** | `Text` | `Body` | Legislative bill text or invoice (rare, 3 training instances). |
| **sidebar** | `Text` | `Sidebar` | Detached aside content. |
| **weather forecast** | `Table` | `Sidebar` | Grid-style structured data region. Per the Visual Grid Rule, 2D-aligned data is a `Table` regardless of domain. |

### Lists and Examination

| M6Doc Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **ordered list** / **unordered list** | `Text` | `ListItem` | M6Doc annotates the container block, not individual items. |
| **first-level question number** | `Text` | `Label` | Top-level exam question label. |
| **second-level question number** | `Text` | `Label` | Sub-question label. |
| **third-level question number** | `Text` | `Label` | Sub-sub-question label. |
| **other question number** | `Text` | `Label` | Question labels not fitting the three levels. |
| **answer** | `Text` | `Value` / `Body` | Exam answer content. |
| **option** | `Text` | `ListItem` | Multiple-choice option. |
| **matching** | `Table` | `FormGroup` | "Match column A to B" exam regions. |
| **examinee information** | `Table` | `FormGroup` | Name / ID input blocks. |

### Visuals and Captions

| M6Doc Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **figure** | `Image` | `Figure` | Standard image or illustration. |
| **mugshot** | `Image` | `Figure` | Headshot portrait photograph. |
| **teasers** | `Image` | `Figure` | Visual preview thumbnails. |
| **table** | `Table` | `Table` | Standard table. |
| **formula** | `Formula` | `DisplayEquation` | Math regions; also annotated inline per guidelines. |
| **caption** | `Text` | `Caption` | Figure caption. |
| **table caption** | `Text` | `Caption` | Table-specific title. |
| **table note** | `Text` | `TableNote` | Notes appended below tables. |
| **credit** | `Text` | `Credit` | Photo source attribution. |
| **barcode** | `OpticalCode` | `Barcode1D` | Linear machine-readable code. |
| **QR code** | `OpticalCode` | `Barcode2D` | 2D machine-readable code. |

### Navigation and Meta

| M6Doc Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **header** | `Text` | `PageHeader` | Running page head. |
| **footer** | `Text` | `PageFooter` | Running page foot. |
| **page number** | `Text` | `PageNumber` | Explicit page index. |
| **folio** | `Text` | `PageNumber` | Combined page number and publication name. |
| **flag** | `Text` / `Image` | `Title` / `Logo` | Newspaper nameplate (e.g., masthead logotype). Not a literal flag image. |
| **inside** | `Text` | `TOC` | "Inside this issue" section pointer. |
| **catalogue** | `Text` | `TOC` | Table of contents or list of works. |
| **index** | `Text` | `Index` | Back-of-book index. |
| **reference** | `Text` | `BibEntry` | Bibliography or works cited. |
| **footnote** / **endnote** | `Text` | `Footnote` | Explanatory meta-text tied to a specific passage. |
| **marginal note** | `Text` | `Annotation` | Text in margins; user-added or editorial marginalia. |
| **dateline** | `Text` | `Dateline` | Temporal and geographic anchor ("NEW YORK, March 4"). |

### Structures and Form Elements

| M6Doc Class | Visual Primitive | Logical Role | Notes |
| :--- | :--- | :--- | :--- |
| **blank** | `Field` | `Input` | Explicit answer box or input area. |
| **underscore** | `Field` | `Input` | Fill-in-the-blank line. |
| **bracket** | `Structure` | `Input` | Answer bracket region in exams. |
| **sealing line** | `Structure` | -- | Security seal line on exam papers. |

## Reproducibility

### Models

TransDLANet uses a ResNet-101 backbone pretrained on ImageNet. The architecture is based on ISTR, extended with a Transformer encoder (no positional encoding), an adaptive element matching mechanism, and a dynamic interaction decoder. Three shared-parameter MLP branches produce classification scores, bounding box coordinates, and segmentation masks. No model weights are released.

### Algorithms

- **Optimizer:** AdamW
- **Base learning rate:** $2 \times 10^{-5}$, decayed to $2 \times 10^{-6}$ at epoch 250 and $2 \times 10^{-7}$ at epoch 375 (50% and 75% of 500 epochs)
- **Batch size:** not reported
- **Augmentation:** random crop; input scaled so the short side is 704-896 px and the long side is at most 1333 px
- **Anchor ratios (for anchor-based baselines):** [0.0625, 0.125, 0.25, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0], a significant departure from the default [0.5, 1.0, 2.0] required to handle M6Doc's wide aspect ratio diversity
- No training or inference code is provided in the repository.

### Data

- **Total:** 9,080 manually annotated pages; 237,116 instances
- **Split:** 6:1:3 (train 5,448 pages / val 908 pages / test 2,724 pages), stratified to maintain label proportions
- **Annotation format:** MS COCO JSON with polygon segmentation masks (not bounding-box only); detection-only baselines use the bounding box derived from each mask's enclosing rectangle
- **Annotation guideline:** 170+ pages; 47 annotators; final QA by authors. No inter-annotator agreement metric (e.g., Cohen's kappa or Fleiss' kappa) is reported. The supplementary material states that inconsistent annotations in the final check were "within 5%," but this is an informal figure without a formal protocol.
- **License:** CC-BY-NC-ND-4.0 for both code and data; prohibits commercial use and derivative datasets
- **Availability:** Test set is publicly downloadable (OneDrive link in repository). Train/val sets require requesting access through the GitHub repository.

### Evaluation

- **Metrics:** Standard COCO mAP (IoU 0.50:0.95), AP50, AP75, and Recall
- **Benchmark:** M6Doc test set; also DocLayNet and PubLayNet test sets for transfer experiments
- **Baselines:** Evaluated with mmdetection framework; anchor configurations adjusted as noted above
- **No official evaluation scripts** are provided; users must reproduce the COCO metric pipeline independently
- **Annotation ambiguities acknowledged:** table vs. paragraph without borders; whether sub-figures count as one or many instances; list item vs. paragraph without explicit markers. These affect reproducibility of edge cases.

### Hardware

Hardware specifications are not reported in the paper.

## BibTeX

```bibtex
@InProceedings{Cheng_2023_CVPR,
    author    = {Cheng, Hiuyi and Zhang, Peirong and Wu, Sihang and Zhang, Jiaxin and Zhu, Qiyuan and Xie, Zecheng and Li, Jing and Ding, Kai and Jin, Lianwen},
    title     = {M6Doc: A Large-Scale Multi-Format, Multi-Type, Multi-Layout, Multi-Language, Multi-Annotation Category Dataset for Modern Document Layout Analysis},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {15138-15147}
}
```