Distance-Annotated Traffic Perception Question Answering (DTPQA)

📄 Paper & Dataset

• 📜 Paper: Read here
• 📁 Dataset: Access here

This repository contains the code used to generate DTPQA, a Visual Question Answering (VQA) benchmark designed to evaluate the perception capabilities of Vision-Language Models (VLMs) in traffic scenes.
DTPQA consists of two main components:

• DTP-Synth, generated using CARLA
• DTP-Real, generated on top of nuScenes data.

📊 Dataset Overview

• 19,149 image–question pairs
• Real and synthetic images
• 10 data types
• Multiple environments and weather conditions
• Distance annotations for each object in question

Code Structure

This repository provides scripts for both the DTP-Synth and DTP-Real datasets.
See the corresponding directories for detailed instructions on how to run the generation pipelines.

BibTeX

@misc{theodoridis2025descriptordistanceannotatedtrafficperception,
      title={Descriptor: Distance-Annotated Traffic Perception Question Answering (DTPQA)}, 
      author={Nikos Theodoridis and Tim Brophy and Reenu Mohandas and Ganesh Sistu and Fiachra Collins and Anthony Scanlan and Ciaran Eising},
      year={2025},
      eprint={2511.13397},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.13397}, 
}