A 3D Reconstruction Benchmark for Asset Inspection

James L. Gray; Nikolai Goncharov; Alexandre Cardaillac; Ryan Griffiths; Jack Naylor; Donald G. Dansereau

cs.CVMar 18, 2026

A 3D Reconstruction Benchmark for Asset Inspection

Authors: James L. Gray, Nikolai Goncharov, Alexandre Cardaillac, Ryan Griffiths, Jack Naylor, Donald G. Dansereau

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

Abstract

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/

Figures & tables

Figure 1: We propose a new benchmark for 3D reconstruction from imagery captured in asset inspection operations. Given user-specified surface soiling, drone trajectories and asset meshes, we generate synthetic data using Blender [ 3 ] that matches the close proximity and high overlap of data captured in industrial asset inspection operations. After adding emulated signal-dependent sensor noise, we benchmark the camera pose and 3D reconstruction accuracy of both end-to-end transformer architectures and traditional structure from motion and multi-view stereo pipelines with ground truth data from our simulator.

Dataset	Captured	Inspection	Surfaces	Depth	Cam.	3D Data	Scenes
ETH3D [ 40 ]	✓	✗	✗	Sparse	✓	✓	25
Urban SaR [ 32 ]	✓	✗	✗	✗	✗	✓	2
PASMVS [ 4 ]	✗	✗	✓	✓	✓	✗	400
BlendedMVS [ 60 ]	✗	✗	✗	✗	✓	✗	110k
DTU [ 19 ]	✓	✗	✗	✗	✓	✓	140
ClaraVid [ 2 ]	✗	✗	✗	✓	✓	✓	40

Table 1: Comparison between our dataset and existing 3D reconstruction and drone imagery datasets. This table compares whether each dataset is captured in an asset inspection style, whether varying levels of surface condition (reflectivity or smudging in our case) are present with the same geometry and what ground truth data is available. Note, Cam. refers to camera parameters. Scenes with an asterisk have variable lighting or environmental conditions.

Figure 2: Renders of the office building , crane and bridge scenes. The office scene was rendered with four surface conditions corresponding to very low, low, medium and high levels of soiling.

			Inputs		Outputs
	Method	Batch	Intrin.	Extrin.	Depth	Normal	PC	Intrin.	Extrin.
E2E	DA3 [ 28 ]	400	$\sim$	$\sim$	✓	✗	✓	✓	✓
	VGGT [ 52 ]	200	✗	✗	✓	✗	✓	✓	✓
	$\pi^{3}$ [ 57 ]	100	✗	$\sim$	✓	✗	✓	✗	✓
SfM/MVS	COLMAP [ 38 , 39 ]	All	$\sim$	$\sim$	✓	✓	✓	✓	✓
SfM	GLOMAP [ 36 ]	All	$\sim$	$\sim$	✗	✗	Sparse	✓	✓

Table 2: Comparison of evaluated methods. ✓ indicates required input and supported output while ✗ indicates unsupported inputs and outputs, and

\sim

refers to optional input. PC refers to dense pointcloud reconstructions and Intrin. and Extrin. are camera intrinsics and extrinsics respectively.

		Office
	Method	V. Low	Low	Med.	High	Crane	Bridge
E2E	DA3 (GT Params.)	0.113	0.114	0.113	0.113	0.207	0.063
	DA3 (Est. Params.)	0.118	0.118	0.118	0.118	0.353	0.075
	VGGT (Est. Params.)	0.200	0.199	0.199	0.201	0.202	0.837
	$\pi^{3}$ (Est. Params.)	0.258	0.260	0.262	0.265	0.625	0.174
MVS	COLMAP (GT Params.)	0.574	0.576	0.570	0.547	1.496	0.755

Table 3: Depth reconstruction accuracy measured as sRMSE (

\downarrow

better) for each scene. Here, Params refers to camera intrinsics and extrinsics. Depth Anything 3 outperforms the the other methods on all scenes except for crane .

Figure 3: Depth maps rendered on the office building (high) scene using self-estimated camera parameters. Each depth map is scaled to match the ground truth range, with near-far corresponding to blue-yellow. Learning-based approaches demonstrate superior performance even in the case of high soiling on reflective scene content, despite recovering less accurate poses. COLMAP is unable to accurately match large regions of non-Lambertian scene content, commonly seen in asset inspection tasks.

		Office
	Method	V. Low	Low	Med.	High	Crane	Bridge
E2E	Depth Anything 3	0.557	0.560	0.555	0.549	0.223	0.195
	VGGT	0.121	0.126	0.137	0.132	0.043	0.036
	$\pi^{3}$	0.538	0.537	0.538	0.537	0.272	0.252
SfM	COLMAP	0.916	0.901	0.934	0.951	0.736	0.205
SfM	GLOMAP	0.763	0.709	0.849	0.804	0.849	0.030

Table 4: Pose accuracy as measured by AUC30 (

\uparrow

better) for each scene. Depth Anything 3 and

\pi^{3}

attain the best results, while VGGT performs poorly.

Figure 4: Qualitative camera pose results on the crane scene. COLMAP, GLOMAP, Depth Anything 3 and

\pi^{3}

performed relatively well at estimating the camera poses. However, VGGT performed poorly.

Figure 5: Point clouds from each method on the bridge scene.

		Office
	Method	V. Low	Low	Med.	High	Crane	Bridge
E2E	DA3 (GT Params.)	0.289	0.290	0.286	0.285	0.307	0.166
	DA3 (Est. Params.)	0.537	0.544	0.540	0.525	0.884	1.043
	VGGT (Est. Params.)	1.070	1.107	0.947	0.922	3.135	3.130
	$\pi^{3}$ (Est. Params.)	0.780	0.784	0.773	0.785	1.891	1.667
MVS	COLMAP (GT Params.)	0.261	0.255	0.260	0.275	0.526	0.079

Table 5: Point cloud Chamfer Distance (m,

\downarrow

better) between the ground truth and estimated point clouds. In contrast to the depth map results, COLMAP with known ground truth camera intrinsics and extrinsics outperforms the other methods.

		Office
	Method	V. Low	Low	Med.	High	Crane	Bridge
E2E	DA3 (GT Params.)	0.243	0.244	0.237	0.236	0.253	0.163
	DA3 (Est. Params.)	0.617	0.625	0.618	0.600	0.744	0.581
	VGGT (Est. Params.)	1.119	1.133	0.913	0.862	1.878	0.966
	$\pi^{3}$ (Est. Params.)	0.356	0.359	0.353	0.351	0.522	0.463
MVS	COLMAP (GT Params.)	0.187	0.186	0.186	0.185	0.575	0.046

Table 6: Point cloud F1-score (m,

\downarrow

better) between the ground truth and estimated point clouds. COLMAP achieves superior performance compared to learning based approaches on all scenes except for the crane, where thin features prove difficult to reconstruct reliably.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: The same frame of the office building scene with the four different soiling settings.

Figure 7: Representative frames from the office building scene.

Figure 8: Representative frames from the crane scene.

Figure 9: Representative frames from the bridge scene.

		Office
	Method	V. Low	Low	Med.	High	Crane	Bridge
E2E	DA3	11.629	11.630	11.220	11.090	6.953	6.090
	VGGT	19.209	18.927	18.221	18.047	8.415	9.776
	$\pi^{3}$	—	—	—	—	—	—
SFM	COLMAP	0.009	0.103	0.134	0.047	0.017	0.144
SFM	GLOMAP	0.051	0.022	0.044	0.011	0.098	0.436

Appendix

Table 7: Mean relative error in focal length(pixels) (%

\downarrow

better). Due to DA3 and VGGT outputing per frame intrinsics we perform a mean. For COLMAP and GLOMAP only a single intrinsics is estimated and compared to. Note that

\pi^{3}

does not output any camera intrinsics.

Method	V. Low	Low	Med.	High	Crane	Bridge
	Office
COLMAP	71.03	69.41	69.16	69.24	79.91	84.82
GLOMAP	76.70	74.93	74.57	73.55	90.34	97.02

Appendix

Table 8: The percentage (%) of registered poses (

\uparrow

better).

Figure 10: Depth maps on the crane scene using self-estimated parameters. Each depth map is scaled to match the ground truth range, with near-far corresponding to blue-yellow. COLMAP is unable to estimate the majority of the frame accurately, whereas transformer-based methods demonstrate good performance in the foreground. However, in the background, performance degrades.

Figure 11: Depth maps on the bridge scene using self-estimated parameters. Each depth map is scaled to match the ground truth range, with near-far corresponding to blue-yellow. COLMAP performs relatively well on the road surface of the bridge and in the foreground, but the suspension cables are estimated to be far too thick. the learning based methods capture the shape of the bridge from this view, but perform poorly in the background.

Figure 12: Camera pose results on the bridge scene. All methods perform poorly.

Figure 13: Camera pose results on the office scene with the very low soiling setting. COLMAP performed best, followed by (in order) GLOMAP, Depth Anything 3,

\pi^{3}

and VGGT.

Figure 14: Point clouds of the office building with very low soiling settings. Depth Anything 3 and COLMAP produce the most accurate point clouds. Note that COLMAP does not place points in the windowed areas, whereas Depth Anything 3 does. $\pi^{3}$ is somewhat misaligned, and VGGT is strongly misaligned.

Figure 15: Point clouds of the crane scene. COLMAP produces very sparse outputs of the crane, mainly reconstructing the background, whereas the other methods mainly reconstruct the crane, with varying degrees of accuracy. Depth Anything 3 produces the most accurate point clouds, while $\pi^{3}$ is somewhat misaligned. VGGT produces unrecognisable output.

3D Reconstruction Benchmarks

See all 15 papers

TerraSky3D: Multi-View Reconstructions of European Landmarks in 4K
Mar 30, 2026Mattia D'Urso, Yuxi Hu, Christian Sormann +23D Reconstruction Multi-View 3D Reconstruction
Graz University of Technology · Sony
VisTa3D: A Dataset and Benchmark for Thin Object Reconstruction from Vision, Tactile, and 3D Point Clouds
Aug 21, 2026Shania Guo, Yeongsik Seo, Andrew Fu +73D Reconstruction Benchmarks 3D Reconstruction
Yale Vision Laboratory, Yale University, New Haven, CT 06520, USA · GRAB Lab, Yale University, New Haven, CT 06520, USA
3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects
May 11, 2026Zhicheng Liang, Haoyi Yu, Boyan Li +63D Reconstruction Benchmarks 3D Reconstruction
The Chinese University of Hong Kong, Shenzhen · Capital Normal University · University of Southern California

Mar 30, 2026J/K move · Enter open · S save

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

Mattia D'Urso, Yuxi Hu, Christian Sormann, Mattia Rossi+1

Graz University of Technology · Sony

3D Reconstruction Multi-View 3D Reconstruction Large-Scale 3D Reconstruction

Open paper AskPDF arXiv

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.

Abstract

3D Reconstruction Benchmarks

Explore similar work

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

VisTa3D: A Dataset and Benchmark for Thin Object Reconstruction from Vision, Tactile, and 3D Point Clouds

3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects