metadata

license: cc-by-nc-4.0
language:
  - en
size_categories:
  - 10K<n<100K
task_categories:
  - image-to-3d
  - image-to-image
  - object-detection
  - keypoint-detection
tags:
  - nerf
  - aerial
  - uav
  - 6-dof
  - multi-view
  - pose-estimation
  - neural-rendering
  - 3d-reconstruction
  - gps
  - imu
pretty_name: AeroGrid100
dataset_info:
  title: >-
    AeroGrid100: A Real-World Multi-Pose Aerial Dataset for Implicit Neural
    Scene Reconstruction
  authors:
    - Qingyang Zeng
    - Adyasha Mohanty
  paper: https://im4rob.github.io/attend/papers/7_AeroGrid100_A_Real_World_Mul.pdf
  workshop: RSS 2025 Workshop on Leveraging Implicit Methods in Aerial Autonomy
  bibtex: |
    @inproceedings{zeng2025aerogrid100,
      title     = {AeroGrid100: A Real-World Multi-Pose Aerial Dataset for Implicit Neural Scene Reconstruction},
      author    = {Zeng, Qingyang and Mohanty, Adyasha},
      booktitle = {RSS Workshop on Leveraging Implicit Methods in Aerial Autonomy},
      year      = {2025},
      url       = {https://im4rob.github.io/attend/papers/7_AeroGrid100_A_Real_World_Mul.pdf}
    }

AeroGrid100

AeroGrid100 is a large-scale, structured aerial dataset collected via UAV to support 3D neural scene reconstruction tasks such as NeRF. It consists of 17,100 high-resolution images with accurate 6-DoF camera poses, collected over a 10×10 geospatial grid at 5 altitude levels and multi-angle views per point.

🔗 Access

To access the full dataset, click here to open the Google Drive folder.

🌍 Dataset Overview

Platform: DJI Air 3 drone with wide-angle lens
Region: Urban site in Claremont, California (~0.209 km²)
Image Resolution: 4032 × 2268 (JPEG, 24mm FOV)
Total Images: 17,100
Grid Layout: 10 × 10 spatial points
Altitudes: 20m, 40m, 60m, 80m, 100m
Viewpoints per Altitude: Up to 8 yaw × 5 pitch combinations
Pose Metadata: Provided in JSON (extrinsics, GPS, IMU)

📦 What’s Included

High-resolution aerial images
Per-image pose metadata in NeRF-compatible OpenGL format
Full drone flight log
Scene map and sampling diagrams
Example reconstruction using NeRF

🎯 Key Features

✅ Dense and structured spatial-angular coverage
✅ Real-world variability (lighting, pedestrians, cars, vegetation)
✅ Precise pose annotations from onboard GNSS + IMU
✅ Designed for photorealistic NeRF reconstruction and benchmarking
✅ Supports pose estimation, object detection, keypoint detection, and novel view synthesis

📊 Use Cases

Neural Radiance Fields (NeRF)
View synthesis and novel view generation
Pose estimation and camera localization
Multi-view geometry and reconstruction benchmarks
UAV scene understanding in complex environments

📌 Citation

If you use AeroGrid100 in your research, please cite:

@inproceedings{zeng2025aerogrid100,
  title     = {AeroGrid100: A Real-World Multi-Pose Aerial Dataset for Implicit Neural Scene Reconstruction},
  author    = {Zeng, Qingyang and Mohanty, Adyasha},
  booktitle = {RSS Workshop on Leveraging Implicit Methods in Aerial Autonomy},
  year      = {2025},
  url       = {https://im4rob.github.io/attend/papers/7_AeroGrid100_A_Real_World_Mul.pdf}
}