TerraSky3D: Multi-View Reconstructions of European Landmarks in 4K

Mattia D'Urso¹ · Yuxi Hu¹ · Christian Sormann² · Mattia Rossi² · Friedrich Fraundorfer¹

¹Graz University of Technology · ²Sony

3DMV at CVPR Workshop 2026

   

Left: Sparse 3D reconstruction of the Villalta Castle, Italy. Right: Representative images collected from aerial and ground perspectives, shown with their corresponding geometrically and semantically filtered depth maps.

Examples of 3DGS reconstructions of aerial and ground scenes from TerraSky3D.

Abstract

Despite the growing need for data of more and more sophisticated 3D reconstruction pipelines, we can still observe a scarcity of suitable public datasets. Existing 3D datasets are either low resolution, limited to a small amount of scenes, based on images of varying quality because retrieved from the internet, or limited to specific capturing scenarios.

Motivated by this lack of suitable 3D datasets, we captured TerraSky3D, a high-resolution large-scale 3D reconstruction dataset comprising ~50,000 images divided into 150 ground, aerial, and mixed scenes. The dataset focuses on European landmarks and comes with curated calibration data, camera poses, and depth maps. TerraSky3D tries to answer the need for challenging dataset that can be used to train and evaluate 3D reconstruction-related pipelines.

📊 Training/Test Set Statistics (v1.0)

Metric	Train	/	Test	Description
Scenes	138	/	12	Number of scenes.
Images	44,505	/	3,018	Number of high-resolution (4K) images provided in the dataset.
Pairs	2,671,062	/	43,720	Total stereo/overlapping image pairs captured.
↳ Ground	1,729,147	/	32,718	Pairs captured only from a ground perspective.
↳ Aerial	728,502	/	6,308	Pairs captured only from an aerial perspective.
↳ Mixed	213,413	/	4,694	Pairs captured from both aerial and ground views.

📁 Download & Format

The dataset can be downloaded from Hugging Face 🤗 .

The Hugging Face repo has one .tar per scene: training scenes are in train/ (named by type: G_ ground, A_ aerial, AG_ mixed aerial and ground), test scenes in test/<scene>.tar, and the test-set pairs in test_pairs/.

Use data_viewer.ipynb to generate train_data.json and visualize examples of pairs of images from all the dataset or just mixed aerial and ground pairs.

---

The dataset follows a structured format for seamless integration into SfM and novel view synthesis pipelines:

data/scene/
  ├── colmap/                  # Structure-from-Motion outputs in COLMAP format
  │   └── sparse/
  │       └── 0/               
  │           ├── cameras      # Intrinsic parameters
  │           ├── images       # Extrinsic parameters / poses
  │           └── points3D     # Sparse point cloud
  ├── frames/                  # Extracted video frames
  │   ├── cam_0/
  │   │   └── frame_000000.jpg # Format: cam_i/frame_*.jpg
  │   ├── cam_1/
  │   │   └── frame_000000.jpg
  │   └── ...
  ├── depth/                   # Multi-View Stereo depth outputs
  │   ├── maps/                # Raw depth estimations from APD-MVS
  │   │   ├── cam_0/
  │   │   │   └── frame_000000.h5
  │   │   └── ...
  │   ├── masks_geometric/     # Geometric masks from APD-MVS
  │   │   ├── cam_0/
  │   │   │   └── frame_000000.png
  │   │   └── ...
  │   └── masks_semantic/      # Semantic masks from Mask2Former
  │       ├── cam_0/
  │       │   └── frame_000000.png
  │       └── ...
  └── train_data.json          # Dictionary containing scenes, images, and camera parameters

🎯 Test Pairs (Relative Pose Benchmark)

test_pairs/ contains all 43,720 pairs of the official test set (12 scenes), with ground-truth intrinsics and relative pose, ready for a pairwise feature-matching / relative-pose benchmark:

• test_pairs/scenes/<scene>.txt: pairs of one scene.
• test_pairs/pairs_all.txt: all scenes in one file.
• test_pairs/pairs_all.json: same pairs grouped by scene, plus image size and radial distortion per camera.

Each TXT line is img0 img1 K0(9) K1(9) R(9) t(3). Image paths are relative to the extracted test scenes (<scene>/frames/<cam>/<image>.jpg), R, t map camera 0 to camera 1 (x1 = R x0 + t, COLMAP world-to-camera), and t is unit-norm. See test_pairs/README.md for details.

🙏 Acknowledgments

We thank Davide Casano, Luca Danelutti, Federico Fattori, Alessandro Menafra, Florian Thaler, Runze Yuan, and Stefano Zorzi for their invaluable contributions to the data collection process.

📝 Citation

If you find this dataset or code useful in your research, please consider citing:

@article{durso202Xterrasky3d,
  title={TerraSky3D: Multi-View Reconstructions of European Landmarks in 4K},
  author={D'Urso, Mattia and Hu, Yuxi and Rossi, Mattia and Sormann, Christian and Fraundorfer, Friedrich},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  year={2026}
}