Datasets:

aldenb
/

web-ui-elements-detection

license: mit
task_categories:
  - object-detection
tags:
  - yolo
  - ui-detection
  - web-screenshots
  - browser-security
  - prompt-injection
size_categories:
  - 100K<n<1M

Web UI Elements Detection Dataset

A large-scale dataset of 268K website screenshots with YOLO-format bounding box annotations for UI element detection. Collected from the Tranco Top 1M website list using automated Chrome-based crawling.

Dataset Summary

Split	Images	Labels	Source
Train (crawled)	268,254	258,497	Automated crawling + annotation
Validation	59	59	Hand-annotated
Test	65	65	Hand-annotated

Classes (5)

ID	Class	Description
0	`interactive`	Clickable elements: buttons, links, tabs, toggles
1	`form_input`	Text fields, search boxes, textareas
2	`form_control`	Checkboxes, radio buttons, dropdowns, sliders
3	`text_label`	Headings, paragraphs, captions, tooltips
4	`media`	Images, videos, icons, avatars, thumbnails

These 5 classes were merged from an original 16-class taxonomy to reduce annotation noise and class imbalance.

Format

Standard YOLO format:

Images: .jpg files (inside tar archives for train split)
Labels: .txt files (inside tar archives for train split), each line: class x_center y_center width height (normalized 0-1)
Val/test splits contain individual files

Directory Structure

train/
  images_000.tar, images_001.tar, ...  # Sharded image archives
  labels_000.tar, labels_001.tar, ...  # Sharded label archives
val/
  images/    # 59 hand-annotated screenshots
  labels/    # YOLO-format annotations
test/
  images/    # 65 hand-annotated screenshots
  labels/    # YOLO-format annotations
dataset.yaml # YOLO dataset config

Extracting Train Data

import tarfile
from pathlib import Path

# Extract all shards
for tar_path in sorted(Path("train").glob("images_*.tar")):
    with tarfile.open(tar_path) as tf:
        tf.extractall("train/images")

for tar_path in sorted(Path("train").glob("labels_*.tar")):
    with tarfile.open(tar_path) as tf:
        tf.extractall("train/labels")

Usage with Ultralytics YOLOv8

from ultralytics import YOLO

# After extracting tar archives:
model = YOLO("yolov8n.yaml")
model.train(data="dataset.yaml", epochs=25, imgsz=640)

Collection Methodology

Screenshots were collected by crawling websites from the Tranco Top 1M list using a Chrome-based crawler (Playwright). Each page was rendered at 1280x720 resolution and annotated using a combination of DOM analysis and heuristic rules mapping HTML elements to the 5 UI classes.

Validation and test sets were manually annotated to ensure high-quality ground truth for evaluation.

Citation

If you use this dataset, please cite:

@misc{web-ui-elements-2025,
  title={Web UI Elements Detection Dataset},
  author={Alden Bernstein},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/aldenb/web-ui-elements-detection}}
}

License

MIT