README.md

---
license: cc-by-4.0
language:
- en
- cs
- es
- fr
- de
- pt
task_categories:
- object-detection
- image-segmentation
tags:
- computer-vision
- object-detection
- coco
- webpage
- ui
- layout-analysis
- accessibility
- aria
- document-layout
pretty_name: WebUI-COCO-868
size_categories:
- n<1K
---


# WebUI-COCO-868: Annotated Webpage Screenshots for UI Region & Landmark Detection

## Short description

868 desktop webpage screenshots annotated with COCO bounding boxes over ARIA-style landmark regions and common UI components (cookie dialogs, popovers, captcha, buttons, etc.)

## Dataset details

- **Modality:** image (webpage screenshots)
- **Annotation format:** COCO-style JSON (bounding boxes + class IDs, optional semantic attributes)
- **Number of images:** **868**
- **Label taxonomy:** ARIA landmark–inspired regions + UI-specific roles (see **Label set** below)
- **Languages (in-page text):** multilingual, reflecting the source domains (primarily English/Czech/Spanish/French/German/Portuguese)

## Supported tasks

This dataset is intended primarily for:

- **Object detection** of semantic webpage regions and UI components (bounding boxes).
- Downstream uses include **web layout parsing**, **UI understanding**, and **accessibility-oriented landmark detection**.

## Repository contents

The dataset is organized into splits with images and per-image JSONs, plus merged COCO files per split:

```
├── annotations/
│ ├── train_annotations.json
│ ├── val_annotations.json
│ └── test_annotations.json
├── train/
│ ├── *.png
│ └── annotations/
│ └── *.json
├── val/
│ ├── *.png
│ └── annotations/
│ └── *.json
├── test/
│ ├── *.png
│ └── annotations/
│ └── *.json
└── split_info.txt
```

- `train/`, `val/`, `test/` contain PNG screenshots named like `123.png`.
- `train/annotations/`, `val/annotations/`, `test/annotations/` contain **per-image** annotation JSON files named like `123.json`.
- `annotations/*_annotations.json` are the **merged COCO** annotation files for each split.
- `split_info.txt` documents how the split was produced (and/or which IDs belong to which split).


## Annotation format

Annotations are stored in **COCO-style JSON**. Each annotation contains:

- `bbox` (x, y, width, height)
- `category_id` (class ID)
- (optional) additional semantic attributes, where applicable

Annotators were instructed to label UI components at **semantically meaningful granularity**:
- Composite widgets (e.g., dialogs) are treated as **containers**
- Controls (e.g., buttons) are labeled as **individual entities**

### Label set (categories)

All classes included in this dataset:

- `main`
- `navigation`
- `image`
- `contentinfo`
- `region`
- `generic-button`
- `banner`
- `complementary`
- `preferences-button`
- `popover`
- `close-button`
- `search`
- `cookie-dialog`
- `accept-button`
- `reject-button`
- `form`
- `captcha`

## Dataset creation

### Collection methodology

The dataset was constructed to capture the **visual and semantic structure of real-world webpages**.

Screenshots were collected using a custom automation pipeline built atop **`nodriver`** (a wrapper for the Chrome DevTools Protocol), enabling fine-grained control over headless browser instances and pixel-accurate rendering.

Key collection details:

- Each session launched **Chromium with Skia rendering** to ensure consistent visual fidelity.
- Cookie consent modals and overlays were programmatically dismissed using **text-based heuristics** and **synthetic clicks**.
- Domain-specific handlers were used to bypass obstructive UI (e.g., paywalls/popups) and ensure content visibility.
- Full-page capture used a **scroll-and-stitch** approach:
  - Incremental scrolling with overlapping screenshots
  - Pinned headers/footers detected via pixel similarity to crop/align slices
  - Pages were optionally stitched offline or stored as individual tiles
- Request interception modified headers (e.g., `User-Agent`, `Referer`) and handled retry/abort logic to reduce failures on dynamic pages and bot protections.
- Screenshots were saved in **timestamped directories** per crawl session.

### Source domains

Webpages were collected from a curated set of high-traffic international media domains to ensure diverse regional styles and layouts, including:

- Bloomberg
- Centrum
- El Cronista
- Financial Times
- Folha de S.Paulo
- Frankfurter Allgemeine Zeitung (Germany)
- Handelsblatt
- iRozhlas
- La Nación
- La Tribune
- Les Echos
- Lidovky
- Novinky
- Seznam
- Seznam Zprávy
- The Business Times
- The Straits Times
- The Wall Street Journal
- Wikipedia

## Quality control

Annotation consistency was validated through:

- Automated heuristics (shape validation, area bounds, basic sanity checks)
- Manual review

## Dataset metrics and diversity

### Element distribution and imbalance

The dataset exhibits a long-tail class distribution:

- The most common class, **`region`**, occurs **> 9000** times.
- Rare classes such as **`captcha`** and **`reject-button`** appear **< 25** times.

This reflects real-world UI patterns: structural regions dominate the page, while edge-case elements are sparse.

### Elements per image

- Mean labeled elements per image: **33.6**
- Some images contain **> 100** labeled elements
- Most images include **~7–11 unique UI types**, indicating moderate layout diversity

### Bounding box shape statistics

Bounding boxes cover a broad range of shapes. Most elements are **horizontally elongated** (wider than tall), consistent with common UI layouts (headers, toolbars, banners). The aspect ratio distribution is skewed toward **wide** and **very wide** elements.

Elements can be grouped by aspect ratio into:
- very tall
- tall
- square-ish
- wide
- very wide

Most elements fall into the **very wide** category.

### Spatial layout patterns

UI elements are not uniformly distributed across the screen:
- concentration along typical layout bands (headers, sidebars, central content)
- clustering patterns consistent with real-world desktop webpage layouts

## Intended use

This dataset is intended for **research and development** in:
- UI element detection on real-world webpages
- webpage semantic region detection (ARIA-style landmarks)
- layout analysis and structured understanding of web interfaces

## Limitations

- **Long-tail imbalance:** rare UI elements (e.g., `captcha`, `reject-button`) have few examples.
- **Domain coverage:** the dataset focuses on a curated set of high-traffic media/information sites and may not generalize to all web genres (e-commerce, web apps, dashboards, etc.).
- **Visual-only labels:** annotations are based on screenshots. They do not include DOM structure or computed accessibility tree.

## Legal and ethical considerations

- Screenshots contain content rendered from third-party websites. Ensure your dataset **license and terms** are compatible with redistribution of rendered pages and any embedded media.
- The crawling pipeline included logic to dismiss overlays and bypass obstructive UI to capture visible content. Users should consider whether this affects allowed redistribution and use.
- The dataset may include incidental personal data if present on captured pages (e.g., author names, user prompts, etc.). Review and redact if needed.

## License

The dataset is released under Creative Commons Attribution 4.0 International (CC BY 4.0).

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{webui_coco_868,
  title     = {WebUI-COCO-868: Annotated Webpage Screenshots for UI Region and Landmark Detection},
  author    = {Lukáš Jílek},
  year      = {2026},
  url       = {https://huggingface.co/datasets/jileklu/WebUI-COCO-868}
}

## Contact
For questions, issues, or takedown requests, please contact:
Name: Lukáš Jílek
Email: jileklu2@fit.cvut.cz
Affiliation: Faculty of Information Technology, Czech Technical University in Prague