💻 computer science

The COTe score: A decomposable framework for evaluating Document Layout Analysis models

This paper introduces the Structural Semantic Unit (SSU) labeling approach and the decomposable COTe score to address the limitations of traditional object detection metrics in Document Layout Analysis, offering a more nuanced and robust framework for evaluating model performance and identifying specific failure modes.

Original authors: Jonathan Bourne, Mwiza Simbeye, Ishtar Govia

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jonathan Bourne, Mwiza Simbeye, Ishtar Govia

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to organize a messy library. You have a pile of books, and your job is to sort them into neat piles: "History," "Science," "Fiction," etc.

For years, computer scientists have used a very strict, rigid ruler to judge how well a robot librarian is doing this job. This ruler is called IoU (Intersection over Union). It basically asks: "Does the robot's box around the book perfectly match the box I drew around the book?"

If the robot draws a box that is slightly too big, or if it draws two boxes where there should be one, the ruler says, "You failed!" even if the robot actually read the title correctly and put the book in the right pile.

This paper argues that this "ruler" is broken for documents (like newspapers, contracts, or magazines) because documents aren't just random 3D objects like cars or dogs; they are puzzles made of text that fit together like tiles on a floor.

Here is the paper's new solution, explained simply:

1. The Problem: The "Rigid Ruler" vs. The "Tessellated Floor"

The authors say that printed pages are like a mosaic floor (or a tessellation). Every piece of text fits snugly next to the next piece. There are no gaps, and pieces shouldn't overlap.

The Old Way (IoU/F1): Imagine judging a tile installer by measuring how perfectly their square tile matches a square hole. If the installer uses a slightly different size tile, or if they accidentally put two tiles on top of each other, the old ruler says, "Zero points!" even if the floor looks fine to a human.
The Issue: In the real world, one person might label a paragraph as "one big box," while a robot might label it as "five small lines." The old ruler thinks the robot is terrible because the boxes don't match perfectly. But the meaning is exactly the same!

2. The New Idea: The "Semantic Unit" (SSU)

The authors introduce a new concept called the Structural Semantic Unit (SSU).

Think of an SSU not as a physical box, but as a story.

If you have a newspaper article, the "SSU" is the whole story.
It doesn't matter if the robot breaks the story into 10 tiny boxes or 1 big box. As long as all those boxes belong to the same story and are next to each other, they count as one unit.

This solves the "labeling mismatch" problem. It stops the judge from caring about the size of the box and starts caring about the meaning of the content.

3. The New Score: COTe

Instead of a single "Pass/Fail" grade, the authors created a new score called COTe. It stands for Coverage, Overlap, Trespass, and Excess.

Think of it like a House Inspector checking a renovation job:

Coverage (Did you paint the whole wall?):
- Did the robot find all the text? If it missed a paragraph, your score goes down.
Overlap (Did you paint over yourself?):
- Did the robot draw two boxes on top of the same sentence? This is like painting the same spot twice. It's messy and confusing. The score goes down.
Trespass (Did you paint the neighbor's wall?):
- Did the robot's box for the "Sports" section accidentally bleed into the "Weather" section? This is the most important part. If the robot mixes up the stories, the text becomes gibberish. The score goes down heavily.
Excess (Did you paint the ceiling?):
- Did the robot draw boxes around empty white space or margins? This is a minor penalty.

The Magic Formula:

COTe Score = Coverage - Overlap - Trespass

If the robot covers everything perfectly but accidentally mixes two stories together (Trespass), the score drops. If the robot misses a whole story (Low Coverage), the score drops. This gives a much clearer picture of why a model is failing.

4. Why This Matters (The "Aha!" Moment)

The authors tested this on real newspaper data and compared it to the old "Rigid Ruler" (IoU/F1).

The Old Ruler: Said the robots were terrible (giving them low scores) just because the robots drew boxes differently than the humans did.
The New COTe Score: Said, "Actually, the robots did a great job! They found all the text and kept the stories separate."

In one test, the old ruler said a model was failing with a score of 0.32, while the new COTe score said it was nearly perfect at 1.00. The old ruler was lying because it was too obsessed with the shape of the boxes, not the quality of the work.

5. The Best Part: You Don't Need to Be Perfect

The authors found something surprising: You don't even need to perfectly define these "Semantic Units" (SSUs) to get the benefit.

Even if you just use the old, messy boxes, the COTe score is still 76% better at telling the truth than the old F1 score. It's like having a smart assistant who can look at a messy room and tell you exactly what's wrong ("You forgot to sweep the kitchen" or "You put the sofa in the hallway") without needing a perfect blueprint of the house first.

Summary

Old Way: "Your box is 2 pixels off, so you failed." (Useless for complex documents).
New Way (COTe): "You found all the text, but you mixed up the sports section with the weather. Fix that, and you're doing great." (Useful for real-world apps).

The paper gives us a new tool to stop wasting time on "perfect boxes" and start focusing on clean, readable, and accurate documents. They even released a free tool (a Python library) so anyone can start using this new "House Inspector" for their own document projects.

1. Problem Statement

Document Layout Analysis (DLA) aims to parse printed media into meaningful elements for digitization. However, the field suffers from a critical evaluation gap:

Inappropriate Metrics: DLA models are currently evaluated using general Object Detection metrics (IoU, F1, mAP) designed for 3D spatial projections (e.g., photography). Printed media is natively 2D, functioning as an irregular tessellation where objects do not overlap and empty space is minimized.
Granularity Mismatch & Incommensurability: Traditional metrics are highly sensitive to the granularity of labeling (e.g., line-level vs. paragraph-level). A model that perfectly extracts text but uses a different labeling granularity than the ground truth receives misleadingly low scores.
Semantic Blindness: Standard metrics treat bounding boxes as atomic units, ignoring the semantic structure of the document. They fail to penalize "Trespass" (a prediction crossing into a different semantic region) or "Overlap" (redundant predictions), which are critical failure modes in DLA that degrade downstream OCR and data extraction.
Pragmatic Failure: The disconnect between training data, test data, and the metric leads to "pragmatic failure," where the metric cannot meaningfully interpret model performance, often suggesting poor performance when the semantic extraction is actually correct.

2. Methodology

The authors propose a new framework consisting of two core components: the Structural Semantic Unit (SSU) and the COTe Score.

A. Structural Semantic Unit (SSU)

Instead of labeling physical text regions (like lines or paragraphs) in isolation, the SSU approach groups labels into relational units based on:

Class: The type of element (e.g., text, title).
Structural Unit: The physical layout container (e.g., a specific column).
Semantic Unit: The narrative or logical grouping (e.g., a specific article).
Reading Order: Adjacency in the reading sequence.

An SSU is defined as a contiguous sequence of text sharing the same class, structural unit, and semantic unit. This shifts the evaluation focus from physical boundaries to semantic coherence.

B. The COTe Score (Coverage, Overlap, Trespass, Excess)

The COTe score is a decomposable metric designed for 2D native imagery. It evaluates the relationship between predictions and ground truth SSUs:

Coverage (C): The normalized area of the ground truth covered by predictions. It rewards covering all content. Unlike IoU, it does not require a strict 1-to-1 match; multiple predictions can contribute to covering a single SSU.
Overlap (O): The area where multiple predictions cover the same ground truth region. This penalizes redundant predictions (stacking), which causes repeated text in OCR.
Trespass (T): The area where a prediction covers a ground truth region belonging to a different SSU. This penalizes boundary breaches that merge unrelated text, causing word order errors.
Excess (e): A support metric measuring the area of predictions outside any ground truth SSU (e.g., predicting margins or gutters).

The Formula:
$\text{COTe Score} = C - O - T$

Range: Theoretically $1 - 2n$ to $1$. A score of 1 is perfect. Negative scores indicate severe failure (more error than coverage).
Granularity Robustness: Because the metric relies on the union of coverage within an SSU rather than strict box alignment, it remains robust even if the prediction granularity differs from the ground truth (e.g., paragraph-level prediction vs. line-level ground truth).

3. Key Contributions

SSU Framework: A relational labeling method that resolves incommensurability between different labeling schemas and granularities by focusing on semantic structure.
COTe Metric: A decomposable, 2D-native evaluation metric that explicitly penalizes semantic boundary breaches (Trespass) and redundancy (Overlap), providing diagnostic insights beyond a single scalar score.
Open Source Tools: Release of the cotescore Python library for applying the framework and a structurally/semantically labeled update of the NCSE dataset.
Empirical Validation: Evaluation of 5 popular DLA models (DocLayout-YOLO, Heron, PP-DocLayout variants) across 3 datasets (NCSE, HNLA2013, DocLayNet).

4. Results

The authors benchmarked the COTe score against traditional metrics (IoU, F1, mAP) and found:

Granularity Robustness: In a case study with perfect text extraction but mismatched granularity (line vs. paragraph), traditional F1 scores dropped to 0.32, while the COTe score remained high (1.00 and 0.84). The COTe reduced the interpretation-performance gap by up to 76% relative to F1.
Diagnostic Power: The decomposable nature of COTe revealed distinct failure modes:
- DocLayout-YOLO: High coverage but suffered from massive Overlap and Trespass, resulting in negative COTe scores on specific pages despite high IoU.
- Heron: Consistently low Trespass but higher Overlap.
- PP-DocLayout: Varied performance based on model size, with smaller models showing higher Trespass.
Correlation: There was only a moderate Spearman correlation between COTe and traditional metrics (e.g., 0.27 for DocLayNet), indicating that COTe ranks models differently and identifies different "best" performers.
No-SSU Feasibility: Surprisingly, the COTe score retained much of its robustness even when applied without explicit SSU labeling (using raw ground truth), suggesting the "many-to-one" relationship between predictions and ground truth is the primary driver of robustness.

5. Significance

Paradigm Shift: The paper argues for moving DLA evaluation from "Object Detection" (3D projection logic) to "Page Parsing" (2D tessellation logic).
Actionable Insights: By breaking down errors into Coverage, Overlap, and Trespass, practitioners can diagnose specific model weaknesses (e.g., "the model merges columns" vs. "the model misses text") and target improvements in architecture or training data.
Lowering Barriers: The finding that COTe works effectively even without complex SSU labeling makes the framework immediately accessible to the community without requiring massive re-annotation efforts.
Standardization: Provides a path toward resolving the "incommensurability" of different DLA datasets and labeling schemas, allowing for fairer model comparison.

In conclusion, the COTe framework offers a more semantically aware, robust, and diagnostic approach to evaluating Document Layout Analysis, addressing the fundamental mismatch between current metrics and the nature of printed media.