Indian Traffic VQA Dataset

🧭 Overview

Indian Traffic VQA is a real-world Visual Question Answering (VQA) dataset focusing on Indian road traffic signboards.The dataset is designed for training and evaluating Vision-Language Models (VLMs) and VQA systems in the traffic and transportation domain.

This dataset bridges a gap between real-world Indian traffic conditions and machine understanding — ideal for research in autonomous driving, smart city AI, and traffic sign recognition under natural environments.

📦 Dataset Summary

Images: 1,085 real-world traffic signboard images

Questions: 4,341 unique questions

Answers: Short, ground-truth textual responses

Source: All images were collected using a mobile phone in real Indian road environments

Format: .csv file with the following columns:

image_name — name of the image file

question — text-based query

answer — corresponding ground-truth answer. The traffic512final.zip file contains all the 1085 images with 512x512 resolution. There are two .csv files attached. traffic_vqa_1085.csv contains one question and one answer, traffic_vqa_4341.csv contains multiple questions and answers per image. The first .csv file can be used for low resource computational environment. 

🧠 Task Definition

Given an image of a traffic signboard and a related question, the model must predict a short text answer.

Example:

image_name

question

answer

image_0001.jpg

What does this sign indicate?

Speed Limit

image_0002.jpg

What does this sign show?

Stop

image_0003.jpg

Is U-turn allowed here?

No

🧩 Applications

Visual Question Answering (VQA)

Vision-Language Model (VLM) Fine-tuning

Multimodal classification of traffic signs

Dataset for benchmarking model reasoning in domain-specific visual data

🧰 Data Collection Details

Captured in diverse Indian traffic conditions (urban, rural, highways)

Includes varying lighting, occlusions, and view angles

All images are real photographs, not synthetic