barkmin77/DriveQA_Dataset

DriveQA: Passing the Driving Knowledge Test

Dataset Summary

DriveQA is a comprehensive multimodal benchmark that evaluates driving knowledge through text-based and vision-based question-answering tasks. The dataset simulates real-world driving knowledge tests, assessing LLMs and MLLMs on traffic regulations, sign recognition, and right-of-way reasoning.

Supported Tasks

• —Text-based QA: Traffic rules, safety regulations, right-of-way principles
• —Vision-based QA: Traffic sign recognition, intersection scene understanding
• —Multimodal Reasoning: Combined visual and textual reasoning for driving scenarios

Dataset Structure

DriveQA-T (Text-based QA)

• —Samples: 26,143 QA pairs + 1,254 challenging samples
• —Categories: 19 subcategories grouped into 5 major domains (Basic Safety, Lane Rules, Special Cases, Road Signs, Emergencies)
• —Format: Multiple-choice questions with explanations

DriveQA-V (Vision-based QA)

• —Samples: 448K image-text QA pairs
• —Image Sources: CARLA simulator + Mapillary real-world data
• —Coverage: 220 US traffic signs, diverse environmental conditions (weather, lighting, perspective, distance)
• —Task Types: Traffic sign recognition, right-of-way judgment

Usage

Organize the Data Structure

After downloading the dataset files, organize them as follows:

1. 1.Place all .jsonl files in your working directory
2. 2.Extract the image archives to the same directory level:

   tar -xzf Intersections_images.tar.gz
   tar -xzf TrafficSigns_CARLA_images.tar.gz  
   tar -xzf TrafficSigns_Mapillary_images.tar.gz

Your final directory structure should look like:

DriveQA/
├── DriveQA_T.jsonl
├── DriveQA_T_HardSet.jsonl
├── DriveQA_V_Intersections_CARLA.jsonl
├── DriveQA_V_TrafficSigns_CARLA.jsonl
├── DriveQA_V_TrafficSigns_Mapillary.jsonl
├── Intersections_images/
├── TrafficSigns_CARLA_images/
└── TrafficSigns_Mapillary_images/

License and Citation

This language dataset is licensed under CC-BY-NC-SA 4.0. If you use this dataset, please cite our work:

@inproceedings{wei2025driveqa,
        title={Passing the Driving Knowledge Test},
        author={Wei, Maolin and Liu, Wanzhou and Ohn-Bar, Eshed},
        booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
        year={2025}
}

Paper dataset for ICCV 2025 DriveQA: Passing the Driving Knowledge Test.

For more information and updates, please visit our website.