# Reading Order Prediction

> Tracking models, datasets, and methods for determining the logical reading sequence of detected document regions.

- Authors: Hunter Heidenreich
- Published: 2023-03-15
- Updated: 2026-04-02
- Source: Bevaya Labs, <https://labs.bevaya.ai/post/reading-order/>

---


> **Disclaimer:** This page covers models for predicting the *logical reading sequence* of detected regions. For detecting *where* regions are on a page, see the [Layout Page](../layout/). For parsing *table internal structure*, see the [TSR Page](../tsr/).

## Overview

Reading order prediction determines the logical sequence in which detected regions should be read. While it depends on layout detection for region proposals, it is a distinct task with its own models, datasets, and evaluation challenges.

The core difficulty is that reading order is not purely spatial. Multi-column layouts, sidebars, footnotes, and floating elements all require understanding the logical flow of a document rather than just scanning top-to-bottom, left-to-right.

> **Note:** We are actively researching this area. Expect significant updates to this section.

---

## Reading Order: Paradigms

Reading order prediction methods organize around five main paradigms:

1. **Rule-based / Projection:** Partitions the page by recursively finding gaps in horizontal or vertical ink projections. No trainable parameters; fast and deterministic.
   - *Examples:* XY-Cut (classic), XY-Cut++
   - *Pros:* No training data required; CPU-deployable; interpretable.
   - *Cons:* Breaks down where clean projection gaps do not exist (L-shaped regions, multi-column with spanning headers, complex wrap-arounds).

2. **Seq2Seq with Pointer Networks:** An encoder reads all layout elements and a decoder emits an ordered sequence by repeatedly pointing to the next element.
   - *Examples:* LayoutReader, FocalOrder
   - *Pros:* End-to-end trainable; captures global context across all elements.
   - *Cons:* Sequential decoding; output length tied to input size; requires large annotated datasets.

3. **Relation Extraction / Dependency Parsing:** Frames reading order as pairwise precedence prediction; a chain-decoding step resolves pairs into a total or partial order.
   - *Examples:* ROOR, Detect-Order-Construct, UniHDSA, HRDoc/DSPS, Token Path Prediction
   - *Pros:* Does not assume a single total order; handles non-linear and partial sequences naturally; compatible with graph-based decoding.
   - *Cons:* Quadratic complexity in element count; requires post-processing to resolve relation pairs into a valid sequence.

4. **Multimodal End-to-End:** A single model reads the document image and generates text, bounding boxes, and semantic labels in reading order in one pass, with no separate layout detection step.
   - *Examples:* ÉCLAIR, DRGG
   - *Pros:* No upstream OCR or detector required; reading order emerges from generation order.
   - *Cons:* Large model sizes; hallucination risk; complex tables and equations often excluded from current releases.

5. **Action-based / Transition Parsing:** Incrementally builds a document hierarchy tree by predicting a discrete action (e.g., insert, skip, merge, move) at each step over a sequence of segments.
   - *Examples:* MTP, HELD, CED/TRACER, CMM, SEG2ACT
   - *Pros:* Naturally captures hierarchical structure (sections, subsections, ToC); interpretable step-by-step construction.
   - *Cons:* Errors propagate through the action chain; many methods require electronic PDFs for font or layout feature extraction.

## Reading Order: Models

### Models with Code or Weights

| Date | Name | Artifacts | Code | License | Notes |
| :--- | :--- | :--- | :--- | :--- | :--- |
| **2025-03** | **UniHDSA** | None | [GitHub](https://github.com/microsoft/CompHRDoc) | MIT | [Notes](2025-03-unihdsa.md). MSRA. Unified relation prediction framework collapsing page-level and document-level HDSA sub-tasks into a single label space. Multimodal Transformer (ResNet + BERT). REDS 96.7 (text), 91.0 (graphical) on Comp-HRDoc. Extends Detect-Order-Construct. |
| **2024-09** | **ROOR baseline** | None | [GitHub](https://github.com/chongzhangFDU/ROOR) | Apache-2.0 | [Notes](2024-09-roor.md). Global pointer network on LayoutLMv3-large. 93.01 F1 (word-level), 82.38 F1 (segment-level) on ROOR dataset. Also introduces RORE downstream enhancement pipeline. |
| **2024-01** | **Detect-Order-Construct** | None | [GitHub](https://github.com/microsoft/CompHRDoc) | MIT | [Notes](2024-01-detect-order-construct.md). MSRA. Multi-modal transformer relation prediction with separate Detect, Order (chain), and Construct (tree) modules with RoPE. REDS 0.9319 on Comp-HRDoc. |
| **2023-04** | **CED/TRACER** | [Spico197/CatalogExtraction](https://github.com/Spico197/CatalogExtraction) | [GitHub](https://github.com/Spico197/CatalogExtraction) | Apache-2.0 | [Notes](2023-04-ced-tracer.md). ICDAR 2023. Compact 3-layer Chinese RoBERTa (RBT3) shift-reduce parser with Sub-Heading/Sub-Text/Concat/Reduce actions for catalog/ToC tree construction. 82.39% F1 on ChCatExt. Also introduces ChCatExt dataset and Wikipedia pre-training corpus. |
| **2021-08** | **LayoutReader** | [microsoft/layoutreader](https://huggingface.co/microsoft/layoutreader) | [GitHub](https://github.com/microsoft/unilm/tree/master/layoutreader) | MIT (code); Unknown (weights) | [Notes](2021-08-layoutreader.md). LayoutLM encoder + pointer-network decoder. Trained on ReadingBank (500k pages). |
| **2021-05** | **MTP** | [stanfordnlp/pdf-struct](https://github.com/stanfordnlp/pdf-struct) | [GitHub](https://github.com/stanfordnlp/pdf-struct) | Apache-2.0 | [Notes](2021-05-mtp.md). Stanford + Google. NLLP@EMNLP 2021. Multimodal transition parser: five-label action sequence (continuous, consecutive, down, up, omitted) predicted by Random Forest on visual and textual features. Works across four document types (scientific, legal, government). Introduces four annotated corpora. |

### Methods (Paper Only)

| Date | Name | Notes |
| :--- | :--- | :--- |
| **2024-05** | **DLAFormer** | ICDAR 2024. MSRA + USTC. Deformable DETR with unified label space treating graphical detection, text region detection, logical role classification, and reading order as a single relation prediction problem. Type-wise query selection (encoder) + type-wise query initialization (decoder). No text embeddings; requires OCR text-line bounding boxes. REDS 96.6% (text) / 90.0% (graphical) on Comp-HRDoc. Full notes in [Layout](../layout/). |
| **2026-01** | **FocalOrder** | [Notes](2026-01-focalorder.md). Unisound AI + CAS. LayoutLMv3-large (0.4B) trained with Focal Preference Optimization: EMA-based difficulty discovery + difficulty-calibrated pairwise ranking. Addresses "Positional Disparity" failure mode. Edit 0.038 EN / 0.055 ZH on OmniDocBench v1.0; REDS 97.1 (text) / 91.1 (graphical) on Comp-HRDoc. |
| **2025-04** | **XY-Cut++** | [Notes](2025-04-xycut-plusplus.md). Rule-based: pre-mask processing + multi-granularity XY-Cut + cross-modal matching. No trainable parameters; 514 FPS on CPU. 98.8 BLEU-4 on DocBench-100. |
| **2025-02** | **ÉCLAIR** | [Notes](2025-02-eclair.md). NVIDIA. ViT-H + mBART decoder (937M params). Jointly extracts text in reading order, bounding boxes, and semantic classes in a single pass. WER 0.142 on DROBS. |
| **2025-02** | **DRGG** | KIT. ICLR 2025. Plug-and-play relation head added to DETR-family detectors; jointly predicts DLA bounding boxes and full relational graph (spatial + logical). InternImage + RoDLA + DRGG: $mAP\_g@0.5 = 57.6\%$, Sequence AP 56.4% on GraphDoc. Full notes in [Layout](../layout/). |
| **2024-10** | **SEG2ACT** | [Notes](2024-10-seg2act.md). EMNLP 2024. CAS. Generative LLM (Baichuan-7B) with a 3-symbol action vocabulary (+, *, =). Global context stack serialized in the same vocabulary; multi-segment batching for efficiency. Single-pass per-segment action prediction. DocAcc on ChCatExt; TEDS on HierDoc. Code: icip-cas/Seg2Act (Unknown license). |
| **2023-10** | **CMM** | [Notes](2023-10-cmm.md). EMNLP 2023. XY-Cut initial tree construction followed by GRU + GAT per-node encoding over a BFS subtree; Keep/Delete/Move operation prediction per node. Introduces ESGDoc (250 Chinese ESG annual reports). Requires electronic PDFs for font extraction. Code: xnyuwg/cmm (Unknown license). |
| **2023-10** | **Token Path Prediction (TPP)** | [Notes](2023-10-token-path-prediction.md). Fudan + Ant Group. EMNLP 2023. GlobalPointer head on LayoutLMv3/LayoutMask predicting directed token-pair edges; no reading-order assumption. Also introduces FUNSD-r and CORD-r benchmarks reflecting realistic OCR disorder. 80.40 F1 (FUNSD-r), 91.85 F1 (CORD-r) with LayoutLMv3. |
| **2023-05** | **LayoutMask** | [Notes](2023-05-layoutmask.md). Ant Group. ACL 2023. Text+layout-only pre-training: local (in-segment) 1D positions + segment-level 2D positions, with Whole Word Masking, Layout-Aware Masking, and Masked Position Modeling (GIoU loss). 182M params (Base). 92.91 F1 on FUNSD, outperforming all T+L+I base models. |
| **2023-05** | **Sparse Graph Segmentation** | [Notes](2023-05-sparse-graph-ro.md). Google. 267K-param multi-task GCN on a $\beta$-skeleton sparse graph. Predicts column-wise vs. row-wise pattern (node) and paragraph membership (edge); hierarchical cluster-and-sort recovers global order. Language-agnostic; mobile-deployable. NED 0.098 on English eval set. ICDAR 2023. |
| **2022-10** | **ERNIE-Layout** | [Notes](2022-10-ernie-layout.md). Baidu. Findings of EMNLP 2022. Layout-aware serialization + spatial-aware disentangled attention + reading order prediction (ROP) as an explicit pre-training task. 24 A100s, 10M IIT-CDIP pages. 93.12 F1 on FUNSD (large). Code + weights: Apache-2.0 (PaddleNLP). |
| **2023-10** | **DSG** | ICDM 2023. ETH Zurich / LMU. Replaces DocParser's heuristic relation rules with LSTM-based prediction + grammar constraints; end-to-end trainable hierarchical parsing. Introduces E-Periodica dataset. Full notes in [Layout](../layout/). |
| **2019-11** | **DocParser** | AAAI 2021. ETH Zurich / LMU. Mask R-CNN detection + rule-based reading-order and parent-child relation classification. Direct predecessor to DSG, HRDoc, and Detect-Order-Construct. Full notes in [Layout](../layout/). |
| **2023-03** | **HRDoc / DSPS** | USTC + iFLYTEK. AAAI 2023. Hierarchical document structure reconstruction as a three-subtask formulation (classify, parent-find, relation-classify). Multi-modal bidirectional encoder + structure-aware GRU decoder with soft-mask. Micro-STEDS 0.8143 (simple) / 0.6903 (hard). Direct predecessor to Detect-Order-Construct. Full notes in [Layout](../layout/). |
| **2021-06** | **ROPE** | [Notes](2021-06-rope.md). Google. Reading Order Equivariant Positional Encoding: assigns relative reading order codes to GCN neighbor messages before aggregation. Plugs into any GCN pipeline. Up to +8.4 F1 on word labeling, +2.5 F1 on FUNSD. ACL-IJCNLP 2021. |
| **2021-05** | **HELD** | [Notes](2021-05-held.md). JCST 2022. BiLSTM/CNN PUT-or-SKIP classifier predicting insertion positions in a rightmost-branch tree. Variable-depth hierarchy from long documents with error-tolerance training via simulated misplacements. Introduces stricter path-correctness metric. No code or data released. |

## Reading Order: Datasets

Datasets are grouped by the most permissive use permitted: **commercial**, **research / non-commercial**, and **not available / restricted**.

### Commercial Use

No reading order datasets with fully permissive licenses (CC0, CDLA-Permissive, MIT, Apache-2.0) are currently confirmed. Datasets in this space tend to carry attribution requirements (CC-BY) or non-commercial clauses, or have undeclared licenses.

### Research / Non-Commercial

| Date | Name | Pages | Domain | Key Contribution | License | Notes |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **2025-11** | **PharmaShip** | 161 docs | Chinese pharmaceutical shipping (scanned) | ISDR reading order graphs + SER + entity linking on domain-specific long-form documents | CC-BY-NC-4.0 | [Notes](2025-12-pharmaship.md). Xi'an Jiaotong University. ~108 RO links/doc vs. ~55 for ROOR. Entity-centric annotation decouples semantics from layout blocks. |
| **2024-09** | **ROOR** | 199 samples | Scanned VrDs (EC-FUNSD-based) | First dataset with ISDR-style reading order as directed acyclic relation pairs; 23.76% non-linear segments | CC-BY-4.0 | [Notes](2024-09-roor.md). Fudan + Ant Group. 10,662 segments, 10,967 annotated ISDR relation pairs. Builds on EC-FUNSD. |
| **2023-10** | **DocTrack** | 539 docs | VRDs (forms, structured docs, infographics) | First eye-tracking-grounded reading order benchmark; human gaze trajectories mapped to OCR bounding box sequences | Research-only (Apache-2.0 header conflicts with non-commercial clause in README; treat as non-commercial until clarified) | [Notes](2023-10-doctrack.md). 39,671 semantic entities; 86,054 tokens. Three subsets: WEAK (FUNSD), STRUCTURED (SeaBill), INFOGRAPH (InfographicVQA). Simple Z-ORDER heuristic often outperforms true eye-tracking order on downstream tasks. |
| **2021-08** | **ReadingBank** | 500k | Diverse (born-digital) | First large-scale reading order benchmark | Research-only (README prohibits redistribution despite Apache-2.0 file) | [Notes](2021-08-layoutreader.md). Auto-extracted from DocX XML metadata via color-based watermarking. English only. |

### Not Available / Restricted

Described in a publication but not publicly downloadable.

| Date | Name | Pages | Domain | Key Contribution | License | Notes |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **2025-02** | **GraphDoc** | 80K images | Mixed (financial, scientific, legal, manuals, patents, government) | Extends DocLayNet with 4.13M relation annotations across 8 categories: Sequence (reading order), Parent, Child, Reference, Up, Down, Left, Right. Only dataset combining spatial and logical graph relations at paragraph level with non-textual elements | Unknown (code: MIT; no dataset license declared; CDLA-Permissive-1.0 inheritance from DocLayNet unconfirmed by authors) | [Notes](2025-02-graphdoc.md). KIT. ICLR 2025. Dataset promised on GitHub but no explicit license statement. Sequence AP = 56.4% with DRGG baseline. |
| **2025-11** | **MosaicDoc** | 72.3K images | Bilingual (ZH/EN) newspapers + magazines (196 publishers) | Block- and line-level reading order on non-Manhattan multi-column layouts; bilingual multi-task (VQA, OCR, ROP, localization) | Not released (code: Apache-2.0) | [Notes](2025-11-mosaicdoc.md). AAAI 2026. Promised open release in Nov 2025; no download available as of Apr 2026. |
| **2023-12** | **M5HisDoc** | 8,000 images (4K regular + 4K hard) | Chinese historical documents (Han to Qing dynasty; 20+ layouts; 20+ doc types) | First multi-style Chinese historical document benchmark; 5 tasks including reading order; Regular + Hard (rotation/distortion/resolution) splits | CC-BY-NC-ND-4.0 (application required) | [Notes](2023-12-m5hisdoc.md). NeurIPS 2023 D&B. SCUT. 16,151 character categories; zero-shot val/test splits. Simple right-to-left heuristic outperforms LayoutReader on this domain. |
| **2020-04** | **HJDataset** | 2,271 scans | Historical Japanese documents (1953 Who's Who) | Layout, hierarchy, and reading order annotations for seven element types; rule-derived right-to-left RO with irregular-case handling | Apache-2.0 (annotations) / Restricted (images require request) | [Notes](2020-04-hjdataset.md). 259,616 annotations. Semi-automatic pipeline: RLSA + CCL segmentation, NASNet Mobile classifier, rule-based RO. No dedicated ROP model trained or evaluated. |

## Reading Order: Metrics

| Metric | What it measures | Notes |
| :--- | :--- | :--- |
| **BLEU** (Page-level) | N-gram overlap between predicted and ground-truth reading order sequence | Used by LayoutReader on ReadingBank. |
| **ARD** (Average Relative Distance) | Positional displacement between predicted and ground-truth element positions; penalty $n$ for omitted elements | Used by LayoutReader and M5HisDoc. Lower is better. |
| **F1 over relation pairs** | Precision and recall over ISDR directed pairs $(i, j)$ | Introduced by ROOR. Replaces BLEU/ARD for the relation-extraction formulation. |
| **REDS** (Reading Edit Distance Score) | Edit distance between predicted and ground-truth reading order sequences, evaluated per group | Primary metric for Comp-HRDoc (Detect-Order-Construct). Lower is better. |
| **NED over block sequences** | Normalized Levenshtein distance between predicted and ground-truth block sequences | Used by OmniDocBench. Lower is better. |
| **Kendall's Tau** | Rank correlation: concordant minus discordant pairs, normalized | Used in some document ordering papers as an alternative to BLEU/ARD. Correlates with human ratings. |
| **Perfect Match Rate** | Fraction of pages where predicted ordering exactly matches ground truth | Used alongside BLEU/ARD as a strict binary measure. |

## Reading Order: Benchmarks

| Date | Name | Size | Scope | Primary Metric | License | Notes |
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
| **2025-04** | **DocBench-100** | 100 docs | Block-level reading order; complex and regular layout subsets | BLEU-4 | Unknown | [Notes](2025-04-xycut-plusplus.md). Introduced with XY-Cut++. Emphasizes multi-column and domain-specific complex layouts underrepresented in existing datasets. |
| **2025-02** | **DROBS** | 789 pages | Joint reading order, bounding boxes, and semantic classes | Counting F1, WER, BLEU | Unknown | [Notes](2025-02-eclair.md). Introduced with ÉCLAIR. Human-annotated from magazines, books, Common Crawl. Tables and equations masked in current release. |
| **2024-12** | **OmniDocBench** | 981 pages | Full document parsing: OCR, layout, table, formula, reading order | NED over block sequences (RO); CDM (formula); TEDS (table) | Apache-2.0 | [Notes](2024-12-omnidocbench.md). CVPR 2025. Nine document types; attribute-level evaluation by language, rotation, background. |
| **2024-09** | **READoc** | 3,576 docs | End-to-end DSE (PDF to Markdown); reading order via KTDS | KTDS (reading order); EDS (text); TEDS (table/ToC) | Unknown | [Notes](2024-09-readoc.md). ACL Findings 2025. Sources: arXiv, GitHub, Zenodo. 16 systems evaluated. |
| **2024-01** | **Comp-HRDoc** | 1,500 docs (1K train / 500 test; HRDoc-Hard extension) | Hierarchical doc structure: page object detection, reading order, TOC, full reconstruction | REDS (reading order); segmentation mAP; Semantic-TEDS | MIT | [Notes](2024-01-detect-order-construct.md). MSRA. First benchmark to evaluate all four subtasks simultaneously. |

---

## To Investigate

Papers and resources identified as likely relevant but not yet fully reviewed.

### Methods

| arXiv | Title | Why Relevant |
| :--- | :--- | :--- |
| `10.18653-v1-2024.emnlp-main.65` | **DocHieNet** (Xing et al., EMNLP 2024) | Document hierarchy parsing with DHFormer (GeoLayoutLM + encoder-decoder); reading order implicit in the hierarchy. Alibaba + Zhejiang. Already in Layout index. No arXiv. |
| -- | **MLARP** (Qiao et al., Pattern Recognition 2024) | Multi-modal layout-aware relation prediction for ROP. Harmonizes visual, semantic, and positional features; layout-aware position embedding handles varying reading habits. No arXiv found; DOI: 10.1016/j.patcog.2024.000657. |
| -- | **XYLayoutLM** (Gu et al., CVPR 2022) | Introduces Augmented XY Cut heuristic to generate multiple total-ordered reading sequences, explicitly acknowledging single-permutation limitation. Used as baseline in M5HisDoc ROP evaluation. |
| -- | **Layout2Pos** (Nguyen et al., TPDL 2024) | Positional encoding injection for reading order; layout-only features; argues spatial layout alone is sufficient to infer reading order without textual content. No arXiv; DOI: 10.1007/978-3-031-72437-4_1. Identified via Giovannini & Marinai survey Tab. 3. |
| -- | **Gao et al. (2013)** - "Newspaper article reconstruction using ant colony optimization and bipartite graph" (Applied Soft Computing 2013) | TSP / ant-colony formulation for reading order recovery across newspaper article blocks. Identified via Giovannini & Marinai survey graph-based ROP category. DOI: 10.1016/j.asoc.2012.07.012. |
| -- | **Quirós & Vidal (2022)** - "Reading order detection on handwritten documents" (Neural Computing and Applications 2022) | Reading order as a learned sorting problem over text region pairs; applied to handwritten documents. No arXiv; DOI: 10.1007/s00521-021-06227-9. |
| -- | **Li et al. (2020)** - "End-to-end OCR text re-organization sequence learning" (ECCV 2020) | GNN + pointer-network approach for reordering OCR text blocks; early deep learning predecessor to LayoutReader. No arXiv. |
| -- | **Clausner et al. (2013)** - "The significance of reading order in document recognition and its evaluation" (ICDAR 2013) | Foundational work defining reading order evaluation methodology; notes single permutation is insufficient and proposes tree structures of ordered/unordered groups. Cited by ROOR as direct theoretical precursor. |
| -- | **Ferilli et al. (2014)** - "Abstract argumentation for reading order detection" (DocEng 2014) | Unsupervised strategy based on abstract argumentation and empirical assumptions about human reading behavior. Historical ML-era method. |
| -- | **Ceci et al. (2007)** - "A data mining approach to reading order detection" (ICDAR 2007) | Early statistical method modeling reading order as multiple disjoint chains; precursor to understanding non-linear reading order. |
| -- | **Aiello & Smeulders (2003/2004)** - "Bidimensional relations for reading order detection" | Foundational rule-based ROP using qualitative rectangle relations; introduced the relational framing that ROOR formalizes two decades later. |
| -- | **Meunier (2005)** - "Optimized XY-Cut for Determining a Page Reading Order" (ICDAR 2005) | The classic XY-Cut algorithm; foundational rule-based baseline that XY-Cut++ directly extends. DOI: 10.1109/ICDAR.2005.182. |
| `2101.12741` | **Post-OCR Paragraph Recognition by Graph Convolutional Networks** (Wang, Fujii & Popat, WACV 2022) | GCN-based grouping and ordering of OCR output into paragraphs. Same Google authors as ROPE; cited by Sparse Graph Segmentation. Reading order implicit in the paragraph assembly step. |
| -- | **Wu, Chou & Chang (2008)** - "A Machine-Learning Approach for Analyzing Document Layout Structures with Two Reading Orders" (Pattern Recognition 2008) | Early ML method explicitly addressing dual reading order in two-column and complex layouts. Cited by FocalOrder as a historical predecessor. No arXiv. DOI: 10.1016/j.patcog.2008.03.009. |
| -- | **MTD** (Hu et al., ICPR 2022) - "Multimodal Tree Decoder for Table of Contents Extraction in Document Images" | Multimodal tree decoder for ToC extraction; introduces HierDoc benchmark dataset. Same group as HRDoc (AAAI 2023) but earlier and distinct paper. Primary baseline for CMM and SEG2ACT. No arXiv found. |
| -- | **Bentabet et al. (ICDAR 2019)** - "Table-of-Contents Generation on Contemporary Documents" | LSTM-based multi-stage ToC generation pipeline; one of the earliest deep learning approaches for ToC/hierarchy extraction; baseline cited in CMM and SEG2ACT. No arXiv found. DOI: 10.1109/ICDAR.2019.00025. |
| `1709.00770` | **Rahman & Finin (BDCAT 2017)** - "Deep Understanding of a Document's Structure" | RNN-based classifier assigning logical structure depth labels (section, subsection, etc.) to document blocks; direct predecessor compared against in HELD and SEG2ACT. |

### Benchmarks / Competitions

| Reference | Title | Why Relevant |
| :--- | :--- | :--- |
| ICDAR 2024 | **RDTAG Task B** (ICDAR 2024) | Competition: "Prediction of Reading Order" for text spotted through AR glasses. 33 registered teams. [Site](https://ilocr.iiit.ac.in/icdar_2024_rdtag/). |

### Metrics

| Metric | Source | Notes |
| :--- | :--- | :--- |
| **Reading-order-independent IE metrics** | Villanova-Aparisi et al., ICDAR 2024 (`2404.18664`) | Metrics for handwritten IE that decouple transcription accuracy from reading order errors. |

---

## Related Pages

- **[Document Layout Analysis](../layout/):** Models and datasets for detecting and classifying regions on a document page.
- **[Table Structure Recognition](../tsr/):** Models and datasets for parsing the internal grid of detected tables.
- **[OCR](../ocr/):** Text recognition pipelines that often run downstream of layout detection and reading order prediction.
- **[Document Understanding](../document-understanding/):** End-to-end systems combining layout detection, structure recognition, and content extraction.



<nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil><nil>