chandrabhuma/Screenspot5G_VQA

📱 Screenspot5G VQA Dataset

🧾 Dataset Details

📌 Dataset Description

  Screenspot5G_VQA is a visual question answering (VQA) dataset for mobile screenshot understanding.
  It is designed to evaluate a model’s ability to reason over real smartphone screen content, including UI elements, icons, layout structure, and visible text.
  All images were captured using a real 5G smartphone, ensuring realistic visual characteristics such as screen resolution, font rendering, and UI density.
  Device: OnePlus Nord CE 2 Lite 5G
  Model Number: CPH2381
  Language: English
  License: CC BY 4.0

📊 Dataset Statistics Split Samples Size Train 1,177 ~33.3 MB Test 295 ~8.7 MB Total 1,472 ~42 MB 👥 Contributors

  🎓 Faculty

      👨‍🏫 Dr. B. Chandra Mohan
      Professor, Dept. of ECE
      Bapatla Engineering College, Bapatla
      
      👨‍🏫 Sri K. Sri Harsha
      Assistant Professor, Dept. of ECE
      Bapatla Engineering College, Bapatla
      
      👨‍🏫 Dr. P. Vinod Babu
      Associate Professor, Dept. of ECE
      Bapatla Engineering College, Bapatla
  
  🎓 Students

      👨‍🎓 Yarramsetty Sindhu
      Undergraduate Student, Dept. of ECE
      Bapatla Engineering College, Bapatla
      
      👨‍🎓 Vasipalli Prasanna
      Undergraduate Student, Dept. of ECE
      Bapatla Engineering College, Bapatla
      
      👨‍🎓 Pilli Harsha Vardhan
      Undergraduate Student, Dept. of ECE
      Bapatla Engineering College, Bapatla
      
      👨‍🎓 Thulava Vamsi
      Undergraduate Student, Dept. of ECE
      Bapatla Engineering College, Bapatla          

🎯 Uses ✅ Direct Use

  This dataset is suitable for:
  
  📱 Mobile UI understanding
  
  👁️ Screenshot-based VQA
  
  🧠 Vision–Language Model (VLM) evaluation
  
  ♿ Accessibility and assistive technologies
  
  🔍 UI element reasoning and screen comprehension

🚫 Out-of-Scope Use

  ### 🚫 Out-of-Scope Use

- 🔒 **Privacy-invasive monitoring**
- 
- ⚠️ **Real-time automated decision-making without human oversight**
- 
- 🔠 **OCR-only benchmarking** (the dataset emphasizes reasoning, not just text extraction)

🗂️ Dataset Structure

  Each sample in the dataset contains the following fields:

- 🆔 **id** *(int64)*  
  Unique sample identifier

- 🖼️ **image_name** *(string)*  
  Filename of the captured screenshot

- ❓ **question** *(string)*  
  Natural language question referring to the screenshot

- ✅ **answer** *(string)*  
  Ground-truth answer corresponding to the question

- 📷 **image** *(image)*  
  Screenshot image used for visual reasoning

The dataset is provided with 🧪 train/test splits to support reproducible evaluation.

🏗️ Dataset Creation

🎯 Curation Rationale

Modern Vision–Language Models (VLMs) often struggle with mobile screen understanding due to:

- 🧩 **Dense UI layouts**
- 🔍 **Small icons and fine-grained fonts**
- 🔤🖼️ **Mixed visual–textual semantics**

Screenspot5G_VQA addresses this gap by using 📱 real-device screenshots, rather than synthetic UI renders, enabling more realistic evaluation of mobile screen understanding.

📥 Source Data 🔧 Data Collection and Processing

  - 📱 **Screenshots captured manually** from a physical smartphone

- ❓ **Questions designed to test:**
  - 🧩 **UI comprehension**
  - 🔤 **Text understanding**
  - 👁️🧠 **Visual grounding and reasoning**

- 🤗 **Dataset formatted for Hugging Face compatibility**

👤 Who are the source data producers?

The dataset was created and annotated by the listed contributors using a personal mobile device.
No automated web scraping or third-party datasets were used.

🏷️ Annotations ✍️ Annotation Process

  - ✍️ **Questions and answers were manually authored**
  - 🎯 **Each question targets a visible element or semantic property of the screen**
  - 🔢 **Single-answer VQA format**

👥 Who are the annotators?

    The contributors listed above performed the annotation and validation.

🔐 Personal and Sensitive Information

    The dataset does not intentionally include personal, private, or sensitive information.
    Screens were curated to avoid identifiable personal data.

⚖️ Bias, Risks, and Limitations

  - 📱 **Screenshots are limited to a single smartphone model**
  - 🤖 **UI design reflects a specific Android ecosystem**
  - 🌐 **Language coverage is limited to English**

📌 Recommendations

  - 🔄 **Combine with datasets from other devices** to improve model generalization
  - 🧠🔤 **Use alongside OCR benchmarks** for comprehensive screen understanding evaluation

📚 Citation

If you use this dataset, please cite: B. Chandra Mohan, K. Sri Harsha, P. Vinod Babu, Y. Sindhu, V. Prasanna, P. Harsha Vardhan, and T. Vamsi, “Screenspot5G_VQA: A Visual Question Answering Dataset for Mobile Screenshot Understanding,” Bapatla Engineering College, Dept. of Electronics and Communication Engineering, Bapatla, India, 2026.