Roadscapes Dataset

A multitask multimodal dataset for autonomous driving scene understanding in diverse Indian road environments.

Overview

Roadscapes is a comprehensive dataset designed to advance research in visual scene understanding for autonomous driving systems, with a particular focus on unstructured and diverse driving environments typical of Indian roads. The dataset consists of 8,981 annotated images capturing various road scenarios, accompanied by question-answer pairs for multiple vision-language tasks.

Key Features

• 8,981 images from diverse Indian driving environments
• Manually verified bounding boxes for object detection in LabelMe JSON format
• Segmentation masks
• Scene attributes derived through rule-based heuristics
• Question-answer pairs for object grounding, visual reasoning, and scene understanding
• Diverse environments: urban, rural, highways, village paths, congested city streets
• Daytime and nighttime captures
• All images anonymized — faces and license plates blurred, no GPS/geotag data

Dataset Statistics

Split	Daytime	Nighttime	Total
Train	5,519	1,989	7,508
Test	1,277	196	1,473
Total	6,796	2,185	8,981

• Train/Test split: ~84% / 16%

Dataset Structure

roadscapes_data/
├── image_data/
│   ├── test/
│   │   ├── Sequence_Day_n/
│   │   │   ├── image.jpg
│   │   │   └── image.json      ← LabelMe format
│   │   └── Sequence_Night_n/
│   └── train/
│       ├── Sequence_Day_n/
│       │   ├── image.jpg
│       │   └── image.json      ← LabelMe format
│       └── Sequence_Night_n/
├── segmentation_masks/
│   ├── test/
│   └── train/
├── vqa_train.csv               ← QA annotations
├── vqa_test.csv                ← QA annotations
└── README.md

Annotation Format

Object Detection — Each image has a paired JSON file in LabelMe format:

{
  "version": "5.0.0",
  "flags": {},
  "shapes": [
    {
      "label": "car",
      "points": [[x1, y1], [x2, y2]],
      "shape_type": "rectangle",
      "flags": {}
    }
  ],
  "imagePath": "image.jpg",
  "imageHeight": 720,
  "imageWidth": 1280
}

VQA Annotations — Stored as CSV files with the following structure:

Column	Description
image_id	Reference to the corresponding image
question	Natural language question about the scene
answer	Ground truth answer
task	Task type (object_counting / object_description / surrounding_description)

Segmentation Masks — Stored as images in the segmentation_masks/ directory, aligned with the corresponding images in image_data/.

Tasks Supported

1. Object Grounding: Locating and identifying objects in road scenes
2. Visual Reasoning: Understanding relationships and interactions between scene elements
3. Scene Understanding: Comprehensive interpretation of driving environments

Capture Setup

Images were captured using a monocular camera across various Indian locations to ensure geographic and environmental diversity, covering different times of day, weather conditions, and road types.

Baseline Model Performance (Accuracy)

Model	Object Counting	Object Description	Surrounding Description
GPT-4o	0.598	0.495	0.701
Paligemma	0.187	0.501	0.485
Phi-3.5	0.667	0.437	0.643
GPT-4o-mini	0.628	0.453	0.645

Privacy & Ethics

• All faces blurred
• All license plates blurred
• No GPS or geotag data included
• No images near sensitive or restricted locations

Citation

If you use Roadscapes in your research, please cite our paper:

@article{iyer2026roadscapesqa,
  title={RoadscapesQA: A Multitask, Multimodal Dataset for Visual Question Answering on Indian Roads},
  author={Iyer, Vijayasri and Rathinagiriswaran, Maahin and S, Jyothikamalesh},
  journal={arXiv preprint arXiv:2602.12877},
  year={2026}
}

License

The dataset is released for non-commercial use only under CC BY-NC 4.0.

Contact

For questions, issues, or collaboration: thisisvij98@gmail.com

Acknowledgments

We thank all contributors involved in data collection, annotation, and verification. This work aims to advance autonomous driving research for challenging and unstructured environments.