Hugging Face's logo

Datasets:
mllmTeam
/
MobileViews

Tasks:
Question Answering
Image-to-Text
Sub-tasks:
task-planning
visual-question-answering
Languages:
English
ArXiv:
Tags:
mobile-ui
user-interfaces
view-hierarchy
android-apps
screenshots
License:
Dataset card Files Community
4
MobileViews / README.md
mllmTeam's picture
mllmTeam
Uploading README.md
0822034 verified
preview code
|
raw
history blame contribute delete
15.8 kB
metadata
language:
  - en
license: mit
datasets:
  - MobileViews
pretty_name: 'MobileViews: A Large-Scale Mobile GUI Dataset'
tags:
  - mobile-ui
  - user-interfaces
  - view-hierarchy
  - android-apps
  - screenshots
task_categories:
  - question-answering
  - image-to-text
task_ids:
  - task-planning
  - visual-question-answering

🚀 MobileViews: A Large-Scale Mobile GUI Dataset

MobileViews is a large-scale dataset designed to support research on mobile agents and mobile user interface (UI) analysis. The first release, MobileViews-600K, includes over 600,000 mobile UI screenshot-view hierarchy (VH) pairs collected from over 20,000 apps on the Google Play Store. This dataset is based on the DroidBot, which we have optimized for large-scale data collection, capturing more comprehensive interaction details while maintaining consistency with DroidBot’s output structure.

📖 Read the Paper — If you’d like to dive deeper into the details of how we construct this dataset, or explore the training experiments we conduct using MobileViews, we encourage you to check out our paper. The paper, along with this dataset, contains everything you need to understand our process and replicate or build upon our findings!

🤗 Table of Contents

🥳 NEWS

  • 2024/11 – New Release! We’ve expanded the dataset to include comprehensive interaction data, now featuring traces and action sequences for 20,000+ apps.
  • 2024/10 – Initial Launch! The MobileViews dataset is live with 600,000+ unique screenshot-VH pairs, deduplicated across 20,000+ Android apps.

🧐 Dataset Overview

Folder Structure

The dataset is organized into two main folders:

  1. MobileViews_Screenshots_ViewHierarchies: Contains the core screenshot-VH pairs in both .zip and .parquet formats. All pairs are globally deduplicated based on their image hash values, ensuring unique UI states across the dataset.
  2. MobileViews_Apps_CompleteTraces: Provides complete interaction traces for individual apps, including states, actions, and transitions.

1. MobileViews_Screenshots_ViewHierarchies

This folder contains 600K screenshot-VH pairs, formatted in both .zip and .parquet formats. Each .zip and .parquet file pair, along with their respective CSV index files, follows a consistent data structure for easy access.

  • Parquets/: .parquet files with binary data for screenshots and JSON content for VHs.
  • Zip_Files/: .zip files with individual .jpg screenshots and .json VH files.
  • Index_CSV/: csv files, like MobileViews_index_0-150000.csv, that map screenshot IDs to their corresponding VH files.

Screenshot-VH File Descriptions

  • MobileViews_0-150000.zip, MobileViews_0-150000.parquet and MobileViews_index_0-150000.csv: This set, containing IDs from 0 to 150,000.
  • MobileViews_150001-291197.zip, MobileViews_150001-291197.parquet and MobileViews_index_150001-291197.csv: This set, containing IDs from 150,001 to 291,197.
  • MobileViews_300000-400000.zip, MobileViews_300000-400000.parquet and MobileViews_index_300000-400000.csv: This set, containing IDs from 300,000 to 400,000.
  • MobileViews_400001-522301.zip, MobileViews_400001-522301.parquet and MobileViews_index_400001-522301.csv: This set, containing IDs from 400,001 to 522,301.

2. MobileViews_Apps_CompleteTraces

The MobileViews_Apps_CompleteTraces folder is a new addition, providing complete interaction traces for apps. Each app trace includes screenshots, VH files, and actions taken within the app, organized as follows:

  • Zip_Files/: Each .zip file represents a batch of app data, named with an index and total screenshot count (e.g., 001_8348.zip : the first batch with 8,348 screenshots).
    • Inside each zip file are 200 subfolders, each named by an app’s package name and containing interaction data for that app. Details on each app’s interaction data structure are provided below.
  • OneExample.zip: A sample app trace illustrating the folder structure and contents for a single app. This zip file includes all files and subfolders needed to explore the sample app trace offline. After downloading and extracting OneExample.zip, you can open index.html in a web browser to interactively view the app trace structure, metadata, and interactions.
  • Index_CSV.zip: This zip file contains all the .csv files listing the app package names and screenshot counts for each corresponding zip file (e.g., 001_8348.csv for 001_8348.zip).

App Trace Folder Structure

Each app’s folder within a zip file includes:

  • states/:
    • Screenshots: .jpg files for each interaction state (e.g., screen_1.jpg).
    • View Hierarchies: .json and .xml files for each screenshot's UI structure (e.g., state_1.json and window_dump_1.xml). The .json format, captured by DroidBot’s internal logic via Android Accessibility Service, and the .xml format, captured via ADB commands, are nearly identical but differ in specific attributes.
  • views/: UI elements related to interactions, such as clicked or scrolled elements for each state transition.
  • actions.csv: A log of interactions within the app, with each action recorded in three columns:
    • from_state: The state before the action.
    • to_state: The state after the action.
    • action: The specific interaction, such as an intent command or touch on a UI element (e.g., <button>, <p>), often with details like the element's alt text and bounding box coordinates.
  • index.html: Open in a browser for an interactive visualization of app traces. Click nodes or edges to view metadata and interaction details for each UI state and transition.
  • utg.js: Supports index.html with structured app trace data, including UI states, transitions, and app metadata.
  • stylesheets/: CSS and JavaScript files for index.html rendering.

🔎 Detailed Metadata and File Descriptions

AppMetadata.csv Columns

The AppMetadata.csv file contains detailed information about each app. Columns include:

Column Description Column Description
title App title installs Number of installs
minInstalls Minimum number of installs realInstalls Real number of installs
score App score (rating) ratings Number of ratings
reviews Number of reviews histogram Rating distribution
price App price free Whether the app is free (True/False)
offersIAP Offers in-app purchases (True/False) inAppProductPrice In-app product price
developer Developer name developerId Developer ID
genre App genre genreId Genre ID
categories App categories contentRating Content rating (e.g., Everyone, Teen)
adSupported App is ad-supported (True/False) containsAds App contains ads (True/False)
released App release date lastUpdatedOn Date of latest update
appId Unique app identifier

Index and Parquet files for MobileViews_Screenshots_ViewHierarchies

  1. Index CSV: Located in MobileViews_Screenshots_ViewHierarchies/Index_CSV, mapping screenshot IDs to their VH files.

    Column Description
    Image File Filename of the screenshot (e.g., 0.jpg)
    JSON File Filename of the VH (e.g., 0.json)

    Example:

    Image File,JSON File
300000.jpg,300000.json
300001.jpg,300001.json

  2. Parquet Columns

    Column Description
    image_content Binary data representing the image file (e.g., screenshot in .jpg format)
    json_content JSON content representing the VH for the corresponding image

    Example Data in Parquet:

    image_content json_content
    Binary image data {"viewHierarchy": {"bounds": [0, 0, 1080, 1920], "viewClass": ...}
    Binary image data {"viewHierarchy": {"bounds": [0, 0, 1080, 1920], "viewClass": ...}
    Binary image data {"viewHierarchy": {"bounds": [0, 0, 1080, 1920], "viewClass": ...}

Index files for MobileViews_Apps_CompeleteTraces

Each index CSV in MobileViews_Apps_CompleteTraces/Index_CSV.zip lists the app package names and screenshot count in the corresponding zip file.

apk_name state_number
com.sportzentral.app 25
com.boostedproductivity.app 23

🌏 How to Download

Download via Hugging Face Python Library

Install the library:

pip install huggingface_hub

Download specific files:

from huggingface_hub import hf_hub_download

# Download specific files
hf_hub_download(repo_id="mllmTeam/MobileViews", filename="MobileViews_Screenshots_ViewHierarchies/Parquets/MobileViews_0-150000.parquet")
hf_hub_download(repo_id="mllmTeam/MobileViews", filename="MobileViews_Screenshots_ViewHierarchies/Zip_Files/MobileViews_0-150000.zip")
hf_hub_download(repo_id="mllmTeam/MobileViews", filename="AppMetadata.csv")

Download the entire repository:

from huggingface_hub

import snapshot_download

# Download the entire repository
snapshot_download(repo_id="mllmTeam/MobileViews")

📋 Usage Guidelines

Using zip and parquet Files in MobileViews_Screenshots_ViewHierarchies

To verify file integrity before using the zip files, follow the steps below.

Zip File Example

# Follow these steps to validate and extract MobileViews_0-150000.zip before using its contents.
zip -T MobileViews_0-150000.zip   # Expected output: test of MobileViews_0-150000.zip OK

# Verify file counts for JSON and JPG files
unzip -l MobileViews_0-150000.zip | grep ".json" | wc -l  # Expected: 150001
unzip -l MobileViews_0-150000.zip | grep ".jpg" | wc -l   # Expected: 150001

# Check file size and SHA256 checksum
du -sh MobileViews_0-150000.zip  # Expected: 23G
sha256sum -c MobileViews_0-150000.zip.sha256  # Expected: MobileViews_0-150000.zip: OK

# Unzip
unzip MobileViews_0-150000.zip

Parquet File Usage

Parquet files are a columnar storage format that allows for efficient data handling, especially for large datasets. Each .parquet file contains binary data for screenshots and JSON data for VHs.

To explore or extract data from .parquet files, use the useparquet.py script provided. This script includes helpful functions, such as check_row_count, save_n_images_and_jsons, and save_all_images_and_jsons, to facilitate quick access to images and VHs.

pip install pyarrow

# Uncomment the function you need and run the helper script to interact with the .parquet file
python path/to/useparquet.py

Mapping Screenshots and States in MobileViews_Apps_CompleteTraces

Each app's folder contains screenshots, states, and actions data. Use our screenshot-state mapping script (MobileViews_Apps_CompleteTraces/screenshot_state_mapping.py) to generate a CSV file mapping screen_id, state_str (a unique hash representing each screenshot), structure_str (a hash focusing on UI structure only), vh_json_id and vh_xml_id.

The script provides two main functions:

  • process_single_folder(single_folder_path) for one apk folder
  • process_multiple_folders(parent_folder_path) for multiple APK folders
# Specify your folder path and uncomment the desired function:
# Run the script
python path/to/screenshot_state_mapping.py

You can also find the example output in MobileViews_Apps_CompleteTraces/OneExample.zip

🌟 Future Updates and Feedback

We’re continuously expanding the MobileViews dataset by crawling more apps and capturing a wider range of interactions. Future releases will bring even more data, making MobileViews an increasingly valuable resource for mobile UI research.

If you have any questions, encounter issues, or want to share feedback, please don’t hesitate to reach out via email or leave a comment. We’re here to help and genuinely appreciate suggestions that could improve the dataset. Your insights are invaluable as we work to make MobileViews as helpful and impactful as possible for the research community.

Thank you for your support and for helping us grow MobileViews!

🔖 Citation

If you use this dataset in your research, please cite our work as follows:

@misc{gao2024mobileviewslargescalemobilegui,
      title={MobileViews: A Large-Scale Mobile GUI Dataset}, 
      author={Longxi Gao and Li Zhang and Shihe Wang and Shangguang Wang and Yuanchun Li and Mengwei Xu},
      year={2024},
      eprint={2409.14337},
      archivePrefix={arXiv},
      primaryClass={cs.HC},
      url={https://arxiv.org/abs/2409.14337}, 
}