README.md

metadata

license: cc-by-sa-4.0
language:
  - zh
  - en
tags:
  - gui
  - agents
pretty_name: FunUI
size_categories:
  - 1K<n<10K

FunUI Benchmark

📖 Introduction

FunUI is a bilingual benchmark designed to fill the gap of comprehensive evaluation datasets in the field of screen understanding.
It encompasses four fundamental tasks and provides a holistic evaluation platform to assess models’ abilities on mobile UI comprehension.

---

✨ Key Features

• Bilingual

  • Includes 2,150 Chinese screens and 9,347 English screens from Android devices.
  • Annotated with about 14k Chinese samples and 18k English samples.
  • The first benchmark that enables systematic evaluation of both Chinese and English screen understanding.

• Comprehensive

  • Covers multiple dimensions of screen understanding:
    • UI Grounding (element localization)
    • UI Referring (element identification)
    • Screen Question Answering
    • Screen Summarization
  • Ranges from spatial grounding and entity recognition to integrated analysis of screen content.

• Diverse

  • Provides QA pairs involving 120+ icons and widgets.
  • Includes complex reasoning questions related to element relations, attributes, arithmetic, and more.
  • Poses greater challenges compared to commonly used OCR-related tasks.

---

📊 Tasks

1. UI Grounding

   • Models are required to localize the target UI element.

2. UI Referring

   • Models identify the specific UI element described in bbox format.

3. Screen Question Answering

   • Models answer diverse questions about screen content.

4. Screen Summarization

   • Models generate summaries of the observed screen.

---

🚀 Applications

• Automated UI comprehension and interaction.
• Development of intelligent assistants and mobile automation.
• Benchmarking multimodal models for screen understanding.

---

📜 Citation

If you use FunUI benchmark in your research, please cite our paper:

@article{202408.2137,
    title = {UI-Hawk: Unleashing the Screen Stream Understanding for GUI Agents},
    author = {Jiwen Zhang and Yaqi Yu and Minghui Liao and Wentao Li and Jihao Wu and Zhongyu Wei},
    doi = {10.20944/preprints202408.2137.v1},
    url = {https://doi.org/10.20944/preprints202408.2137.v1},
    year = 2024,
    month = {August},
    publisher = {Preprints},
    journal = {Preprints}
}