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

Mattia D'Urso; Yuxi Hu; Christian Sormann; Mattia Rossi; Friedrich Fraundorfer

cs.CVMar 30, 2026

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

Authors: Mattia D'Urso, Yuxi Hu, Christian Sormann, Mattia Rossi, Friedrich Fraundorfer

Organizations: Graz University of Technology · Sony

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.

Figures & tables

Panel of Figure 1 — Figure 1: Example scene from TerraSky3D. Left: Sparse reconstruction of the Villalta Castle, Italy. Right: Representative images collected from aerial and ground perspectives, shown with their corresponding semantically filtered depth maps.

Figure 2: Geographical Distribution of Data Collection Sites in TerraSky3D . The dataset includes sites across 11 European countries. Close locations might share the same red pin.

Dataset	Year	#Scenes	Type	Real	HR	Calib.	COLMAP	Depth
MegaDepth [ 17 ]	2018	135	G	✓	✗	✗	✓	COLMAP-MVS
MegaScenes [ 27 ]	2024	100K	G	✓	✗	✗	✓	✗
University-1652 [ 36 ]	2020	1652	M	✗ / ✓	✗	✗	✓	COLMAP-MVS
CVD-SfM [ 16 ]	2025	3	M	✓ / ✓	✓	✗	✓ +GPS	✗
Culture3D [ 35 ]	2025	41	M	✓ / ✓	✓	✓	✓ +RC	✗
AerialMegadepth [ 29 ]	2025	137	M	✗ / ✓	✗	✗	✓	COLMAP-MVS

Table 1: Comparison of MVS and 3D Reconstruction Datasets. An ✗ i under Scenes indicates that the number of scenes is not available. Types are G (Ground) or M (Mixed Air and Ground). ✓ under Real indicates collected imagery, whereas ✗ i denotes computer-generated images; both aerial and ground are indicated when available (e.g., ✗ / ✓ means aerial is computer-generated but ground is real). HR denotes High Resolution (Full HD or above). Calib. indicates camera pre-calibration ( ✓ ) or intrinsic estimation during the reconstruction process ( ✗ ). ✓ i COLMAP indicates geometry estimated with COLMAP, while ✗ i indicates no geometry is available. RC stands for Reality Capture.

Panel of Figure 3 — Figure 3: Example Scene from TerraSky3D. Left: Sparse reconstruction of the Natural History Museum, Vienna, Austria. Right: The first row shows example images, and the second row shows the corresponding semantically filtered depth maps.

Figure 4: Visualization of the Depth Filtering Process. From left to right: original RGB image, raw depth map from APD-MVS, semantic mask, APD-MVS confidence mask, and the filtered depth map. The scene depicted is the Arch of Victory , Madrid, Spain.

Panel of Figure 5 — Figure 5: Example Scene from TerraSky3D. Left: Sparse reconstruction of the Italian Charnel House, Kobarid, Slovenia. Right: Representative images collected from aerial and ground perspectives, shown with their corresponding semantically filtered depth maps.

		AUC@5 $\uparrow$
	Method	Ground	Aerial	Mixed	Mean
Sparse FE	SuperPoint [ 5 ]	53.2	57.9	17.7	42.9
	DISK [ 28 ]	56.1	72.1	29.5	52.6
	S-TREK [ 21 ]	62.0	81.8	36.9	60.2
	DeDoDe-G [ 7 ]	62.7	78.8	35.5	59.0
	RDD [ 3 ]	62.3	74.5	28.7	55.2

Table 2: Relative Pose Estimation Performance on the TerraSky3D Test Set. We report performance in terms of AUC@

5^{\circ}

. The keypoints budget is set to 4096. MD and TS3D denote models trained on MegaDepth and TerraSky3D.

		AUC@5 $\uparrow$
	Imgs	VGGT [ 30 ]	MapAny. [ 11 ]	$\boldsymbol{\pi^{3}}$ [ 31 ]
Graz Church	239	71.4	53.7	0 0.0
Graz Townhall	335	17.9	51.7	26.9
Graz University	382	33.1	26.4	12.0
Munich Frauenkirche	123	68.2	65.2	77.3
Munich Marienplatz	246	56.4	60.0	68.5

Table 3: Absolute Pose Estimation Performance on the TerraSky3D Test Set. We report performance in terms of AUC@

5^{\circ}

. MapAny. stands for MapAnything.

Scene		Type	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Training set	Trieste Unity Square	G	22.21	0.759	0.318
	Vienna Natural History Museum	G	21.92	0.754	0.342
	Udine Abandoned Bridge	A	21.73	0.625	0.326
	Gorizia Military Shrine	M	21.53	0.656	0.398
	Udine Villalta Church	M	23.76	0.746	0.302
Test set	Graz Church	G	22.39	0.801	0.197

Table 4: NVS Benchmark Results on TerraSky3D. Scene types are G (Ground), A (Aerial), and M (Mixed Aerial and Ground).

Panel of Figure 6 — Figure 6: Example Scene from TerraSky3D. Left: Sparse reconstruction of Erto e Casso, Pordenone, Italy. Right: Representative images collected from aerial and ground perspectives, shown with their corresponding semantically filtered depth maps.

				Inliers (%)
	Scene	Type	Pairs	1px	3px	5px	10px
MegaDepth	British Museum	G	4104	39.0	51.3	53.4	54.9
	Florence Cathedral Side	G	4134	47.8	62.3	65.3	67.7
	Lincoln Memorial Statue	G	3195	36.6	56.4	62.7	67.9
	London Bridge	G	2200	42.1	56.1	60.5	65.0
	Milan Cathedral	G	4279	50.6	64.5	67.9	70.8

Table 5: Bidirectional Geometric Consistency on TerraSky3D. We report performance in terms of cumulative inlier percentage under varying pixel error thresholds.

Panel of Figure 7 — Figure 7: Example Scene from TerraSky3D. Left: Sparse reconstruction of the Barcis Dam, Pordenone, Italy. Right: Representative images collected from aerial and ground perspectives, shown with their corresponding semantically filtered depth maps.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 930 papers

A 3D Reconstruction Benchmark for Asset Inspection
Mar 18, 2026James L. Gray, Nikolai Goncharov, Alexandre Cardaillac +33D Reconstruction Benchmarks 3D Reconstruction
Australian Centre for Robotics, School of Aerospace, Mechanical and Mechatronic Engineering, University of Sydney, Sydney, NSW, Australia
Scene Grounding In the Wild
Mar 27, 2026Tamir Cohen, Leo Segre, Shay Shomer-Chai +2Large-Scale 3D Reconstruction 3D Reconstruction
Tel Aviv University · Cornell University
SeasonScapes: Learning Large-scale Re-lightable 3D Landscapes with Seasonal Variation from Sparse Webcams
May 9, 2026Timo Kleger, Qi Ma, Deheng Zhang +2Dynamic Scene Reconstruction Sparse-View 3D Reconstruction
2ETH Zurich · 1INSAIT, Sofia University “St. Kliment Ohridski”

Mar 18, 2026J/K move · Enter open · S save

A 3D Reconstruction Benchmark for Asset Inspection

James L. Gray, Nikolai Goncharov, Alexandre Cardaillac, Ryan Griffiths+2

Australian Centre for Robotics, School of Aerospace, Mechanical and Mechatronic Engineering, University of Sydney, Sydney, NSW, Australia

3D Reconstruction Benchmarks 3D Reconstruction Robotic Inspection

Open paper AskPDF arXiv

Asset management requires accurate 3D models to inform the maintenance, repair, and assessment of buildings, maritime vessels, and other key structures as they age. These downstream applications rely on high-fidelity models produced from aerial surveys in close proximity to the asset, enabling operators to locate and characterise deterioration or damage and plan repairs. Captured images typically have high overlap between adjacent camera poses, sufficient detail at millimetre scale, and challenging visual appearances such as reflections and transparency. However, existing 3D reconstruction datasets lack examples of these conditions, making it difficult to benchmark methods for this task. We present a new dataset with ground truth depth maps, camera poses, and mesh models of three synthetic scenes with simulated inspection trajectories and varying levels of surface condition on non-Lambertian scene content. We evaluate state-of-the-art reconstruction methods on this dataset. Our results demonstrate that current approaches struggle significantly with the dense capture trajectories and complex surface conditions inherent to this domain, exposing a critical scalability gap and pointing toward new research directions for deployable 3D reconstruction in asset inspection. Project page: https://roboticimaging.org/Projects/asset-inspection-dataset/

Abstract

3D Reconstruction

Explore similar work

A 3D Reconstruction Benchmark for Asset Inspection

Scene Grounding In the Wild

SeasonScapes: Learning Large-scale Re-lightable 3D Landscapes with Seasonal Variation from Sparse Webcams