jarvissun/UAVReason_vqa

🚁 UAVReason: VQA / Caption / Generation Annotations

A UAV-native benchmark for aerial visual reasoning, captioning, temporal understanding, and cross-modal generation

📌 News

• Paper: Can Vision-Language Models Think from the Sky? Unifying UAV Reasoning and Generation
• arXiv: arXiv:2604.05377
• Depth data: jarvissun/UAVReason_depth

This dataset is released as part of UAVReason, introduced in the paper above.
Please cite the paper if you use this dataset.

🧭 Overview

UAVReason is designed for studying whether vision-language models can reason from high-altitude UAV-view scenes. Unlike ground-view images, UAV scenes contain tiny and densely distributed objects, repetitive textures, large viewpoint changes, and ambiguous top-down orientations. This repository provides the language-side annotation files of UAVReason, including visual question answering, captioning, and cross-modal generation instructions.

The full UAVReason benchmark aligns RGB imagery, depth maps, semantic segmentation masks, captions, and question-answer pairs in a consistent aerial domain. It supports unified research on:

• Aerial visual question answering
• Scene captioning
• Spatial reasoning
• Temporal reasoning
• Heading-aware reasoning
• Depth-aware multimodal understanding
• Cross-modal generation across RGB, depth, and segmentation

📊 Dataset Highlights

Component	Description
Captioned frames	UAV-view frames with scene-level textual descriptions
VQA pairs	Single-frame and two-frame question-answer pairs
Temporal questions	Two-frame questions for temporal scene reasoning
Generation annotations	JSONL instructions for cross-modal generation tasks
Modalities	RGB, text, depth, semantic segmentation
Domain	UAV / aerial / nadir-view scenes

The associated paper reports 23.6K captioned frames, 273K VQA pairs, including 68.2K two-frame temporal questions, and 188.8K cross-modal generation samples across RGB, depth, and segmentation modalities.

📁 Repository Contents

This repository contains annotation files for UAVReason. Depending on the version, the file tree may include JSONL files for:

UAVReason_vqa/
├── README.md
├── *.jsonl                 # VQA / caption / generation annotations
└── ...

Each JSONL file stores one sample per line. A sample may contain image references, textual prompts, questions, answers, captions, task labels, or modality-condition information. Please inspect the keys of each JSONL file before training, because different tasks may use different fields.

🚀 Quick Start

1. Download

pip install -U huggingface_hub datasets

hf download jarvissun/UAVReason_vqa \
  --repo-type dataset \
  --local-dir ./UAVReason_vqa

2. Inspect JSONL files

import json
from pathlib import Path

root = Path("./UAVReason_vqa")
jsonl_files = sorted(root.rglob("*.jsonl"))
print("JSONL files:")
for p in jsonl_files:
    print(" -", p.relative_to(root))

# Print one example from the first JSONL file
with jsonl_files[0].open("r", encoding="utf-8") as f:
    example = json.loads(f.readline())

print(example.keys())
print(json.dumps(example, indent=2, ensure_ascii=False)[:2000])

3. Minimal VQA-style loading example

import json
from pathlib import Path

jsonl_path = Path("./UAVReason_vqa/your_file.jsonl")

samples = []
with jsonl_path.open("r", encoding="utf-8") as f:
    for line in f:
        item = json.loads(line)
        samples.append(item)

print("Number of samples:", len(samples))
print("Available keys:", samples[0].keys())

🧪 Suggested Tasks

1. Single-frame UAV VQA

Given one UAV-view image and a question, the model predicts the answer. This task evaluates object recognition, counting, attribute understanding, spatial relations, and aerial scene grounding.

2. Two-frame Temporal VQA

Given two UAV-view frames and a question, the model predicts the answer by comparing scene changes across time. This task evaluates temporal reasoning, motion awareness, and consistency across viewpoints.

3. Heading-aware Reasoning

Given a UAV-view scene and a heading-sensitive question, the model reasons about directions and orientation under the top-down view.

4. Captioning

Given one UAV-view image, the model generates a concise scene description that captures objects, layout, and high-level semantics.

5. Cross-modal Generation

Given text and optional dense conditions such as depth or semantic segmentation, the model generates or reconstructs the target modality. This task studies unified aerial understanding and generation.

🔗 Related Repositories

Resource	Link
Paper page	Hugging Face Papers
arXiv	arXiv:2604.05377
Depth data	UAVReason_depth
This dataset	UAVReason_vqa

🧾 Citation

If you use UAVReason in your research, please cite:

@article{sun2026uavreason,
  title={Can Vision-Language Models Think from the Sky? Unifying UAV Reasoning and Generation},
  author={Sun, Jintao and Ding, Gangyi and Di, Donglin and Zhang, Hu and Zheng, Zhedong},
  journal={arXiv preprint arXiv:2604.05377},
  year={2026}
}

📄 License

This dataset is released under the Apache-2.0 License. Users should also respect the terms of any upstream data sources and use the dataset only for lawful research and development purposes.

🙏 Acknowledgement

UAVReason is built for advancing UAV-view multimodal understanding and generation. We hope it can support research on aerial visual reasoning, spatial intelligence, temporal understanding, and geometry-aware multimodal generation.