arxiv: 2604.19596 · v3 · submitted 2026-04-21 · 💻 cs.CV

Recognition: unknown

PC2Model: ISPRS benchmark on 3D point cloud to model registration

Mehdi Maboudi , Said Harb , Jackson Ferrao , Kourosh Khoshelham , Yelda Turkan , Karam Mawas

Authors on Pith no claims yet

Pith reviewed 2026-05-10 03:35 UTC · model grok-4.3

classification 💻 cs.CV

keywords point cloud registration3D model alignmentbenchmark datasetsimulated point cloudsreal-world scansdomain transferLiDAR dataconstruction monitoring

0 comments

The pith

The PC2Model benchmark provides a hybrid dataset of simulated and real point clouds paired with 3D models to train and evaluate registration methods.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes a public benchmark dataset called PC2Model for aligning point clouds to three-dimensional models. Data-driven approaches currently falter on real scans due to sparsity, noise, clutter, and occlusions, limiting their use in construction monitoring, autonomous driving, robotics, and VR/AR. The benchmark combines simulated point clouds, which supply exact ground truth under controlled conditions, with real-world scans that include authentic sensor and environmental artifacts. This hybrid structure is intended to support training across domains and to enable direct analysis of how well models transfer from simulation to reality.

Core claim

The paper introduces the PC2Model benchmark as a publicly available dataset that pairs simulated point clouds with, in some cases, real-world scans and their corresponding 3D models. Simulated portions deliver precise ground truth and controlled conditions while real portions introduce sensor and environmental artefacts, allowing systematic study of domain transfer and robust training of both classical and data-driven registration methods.

What carries the argument

The PC2Model benchmark dataset, whose hybrid design of simulated point clouds combined with real-world scans and 3D models supplies controlled ground truth alongside authentic artefacts for cross-domain evaluation.

If this is right

Both classical and data-driven registration methods can be trained and compared under identical controlled and realistic conditions.
Transferability of learned models from simulation to real scans can be quantified and improved systematically.
Standardized evaluation becomes possible for downstream tasks that require point cloud to model alignment.
Public access allows the community to extend the dataset or test new methods against a common baseline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Methods that succeed on this benchmark could be adapted to improve real-time alignment in robotics or augmented reality systems where ground truth is unavailable.
The same hybrid construction approach might be applied to other 3D vision tasks such as object detection or semantic segmentation to study domain gaps.
If transfer performance improves markedly, the benchmark could reduce the need for large-scale real-world labeled data collection in future registration research.

Load-bearing premise

The hybrid design of simulated point clouds with real-world scans will enable systematic analysis of model transferability from simulated to real-world scenarios and support robust training across domains.

What would settle it

A direct comparison experiment showing that models trained only on the simulated portion achieve no measurable improvement in registration accuracy or robustness when tested on the real-world portion, compared with models trained directly on real data alone.

Figures

Figures reproduced from arXiv: 2604.19596 by Jackson Ferrao, Karam Mawas, Kourosh Khoshelham, Mehdi Maboudi, Said Harb, Yelda Turkan.

**Figure 1.** Figure 1: Point cloud simulation from 3D model and its co-registration. a) 3D model, b) point cloud sampled from 3D model, c) point [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: Occluded regions (yellow) and scanned regions (blue) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Surface deviation from the ground-truth model. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: Mixed-pixel artefacts along object edges, caused by [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: Point density variation across the surface; warmer [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

**Figure 6.** Figure 6: Representative dataset samples spanning seven categories: (a) mechanical objects, (b)furniture, (c) home d [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

**Figure 7.** Figure 7: Category-level LOA of PC2Model objects. LOC increases with higher threshold values for all categories. Simulated categories generally achieve high coverage, indicating that most of the model’s surfaces are well represented by the point cloud. However, categories like furniture show lower coverage at stricter thresholds, mainly due to occlusions. Indoor spaces (real) demonstrate comparatively lower LOC at … view at source ↗

**Figure 8.** Figure 8: Level of Coverage (LOC) distributions across [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗

read the original abstract

Point cloud registration involves aligning one point cloud with another or with a three-dimensional (3D) model, enabling the integration of multimodal data into a unified representation. This is essential in applications such as construction monitoring, autonomous driving, robotics, and virtual or augmented reality (VR/AR). With the increasing accessibility of point cloud acquisition technologies, such as Light Detection and Ranging (LiDAR) and structured light scanning, along with recent advances in deep learning, the research focus has increasingly shifted towards downstream tasks, particularly point cloud-to-model (PC2Model) registration. While data-driven methods aim to automate this process, they struggle with sparsity, noise, clutter, and occlusions in real-world scans, which limit their performance. To address these challenges, this paper introduces the PC2Model benchmark, a publicly available dataset designed to support the training and evaluation of both classical and data-driven methods. Developed under the leadership of ICWG II/Ib, the PC2Model benchmark adopts a hybrid design that combines simulated point clouds with, in some cases, real-world scans and their corresponding 3D models. Simulated data provide precise ground truth and controlled conditions, while real-world data introduce sensor and environmental artefacts. This design supports robust training and evaluation across domains and enables the systematic analysis of model transferability from simulated to real-world scenarios. The dataset is publicly accessible at: https://zenodo.org/records/17581812

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper releases PC2Model, a hybrid simulated-plus-real benchmark for point cloud to model registration, but the key claim that it enables systematic sim-to-real transfer analysis lacks supporting details or controls.

read the letter

The main takeaway is that the authors have created and made public PC2Model, a benchmark dataset pairing simulated point clouds with real-world scans and their 3D models. It targets registration tasks in construction, robotics, and AR/VR where data-driven methods often fail on noise and occlusions. The hybrid setup is the central idea: simulation for clean ground truth and control, real data for sensor artifacts. The dataset is hosted on Zenodo and backed by an ISPRS working group, which helps with visibility and potential adoption as a standard testbed. That public release and the practical focus are the clear positives here. Researchers building or testing registration pipelines now have another option beyond purely synthetic or purely real collections. The design intent makes sense on paper for studying domain gaps. The soft spot is exactly what the stress-test note flags. The abstract asserts that the hybrid structure supports systematic transferability analysis and robust cross-domain training, yet it provides no generation parameters, no statistical comparisons between simulated and real subsets, no noise or occlusion models, and no baseline registration results that quantify any transfer gap. Without those, the utility for isolating domain effects remains an untested assertion rather than a property users can rely on immediately. If the full manuscript adds those controls and some quantitative validation, the paper strengthens considerably. Otherwise the data may still be usable but will require extra work from adopters to verify the claimed benefits. This is for computer vision groups working on 3D registration who need fresh benchmark data for training or evaluation. A reader already active in point cloud methods or looking for ISPRS-aligned resources will get the most out of it. I would send it to peer review. The dataset contribution is concrete enough to merit referee time even if the transferability section needs more evidence.

Referee Report

2 major / 0 minor

Summary. The paper introduces the PC2Model benchmark, a publicly available dataset for point cloud-to-model (PC2Model) registration tasks. It adopts a hybrid design combining simulated point clouds (with precise ground truth and controlled conditions) and, in some cases, real-world scans (introducing sensor and environmental artifacts) along with corresponding 3D models. The benchmark is intended to support training and evaluation of both classical and data-driven methods while enabling systematic analysis of model transferability from simulated to real-world scenarios. The dataset is released via Zenodo.

Significance. If the dataset is accompanied by detailed, reproducible construction protocols, quantitative characterizations, and baseline experiments, it could provide a valuable standardized resource for the 3D computer vision and photogrammetry communities. This would help address persistent challenges in real-world registration such as sparsity, noise, clutter, and occlusions, with direct relevance to applications in construction monitoring, robotics, and autonomous systems.

major comments (2)

[Abstract] Abstract: The central claim that the hybrid design 'supports robust training and evaluation across domains and enables the systematic analysis of model transferability from simulated to real-world scenarios' is asserted without any reported details on data generation procedures (e.g., explicit noise models, occlusion patterns, density variations, or sensor simulation parameters), statistical comparisons between simulated and real subsets, or baseline registration results that quantify domain gaps or transfer performance. This renders the asserted utility an untested design intent rather than a demonstrated property of the released data.
[Abstract] Abstract: No quantitative dataset characteristics (e.g., number of point clouds, average point counts, ground-truth alignment error distributions, or validation metrics) or description of how real-world scans were paired with models are provided. For a benchmark paper, these are load-bearing elements needed to assess reproducibility and suitability for the claimed cross-domain training and transferability analysis.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the PC2Model benchmark paper. We agree that the abstract requires strengthening with quantitative details and data generation descriptions to better support the claims. We will revise the manuscript accordingly.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that the hybrid design 'supports robust training and evaluation across domains and enables the systematic analysis of model transferability from simulated to real-world scenarios' is asserted without any reported details on data generation procedures (e.g., explicit noise models, occlusion patterns, density variations, or sensor simulation parameters), statistical comparisons between simulated and real subsets, or baseline registration results that quantify domain gaps or transfer performance. This renders the asserted utility an untested design intent rather than a demonstrated property of the released data.

Authors: We agree that the abstract presents the utility claim without the supporting details listed, which leaves it as an assertion rather than a demonstrated result. The manuscript describes the hybrid design at a high level and provides the Zenodo release, but does not include the specific quantitative elements or baselines in the abstract or main text. We will revise by adding explicit descriptions of data generation procedures (noise models, occlusion patterns, density variations, sensor parameters), statistical comparisons between subsets, and baseline registration results quantifying domain gaps and transfer performance. The abstract will be updated to summarize these additions and reference the new content. revision: yes
Referee: [Abstract] Abstract: No quantitative dataset characteristics (e.g., number of point clouds, average point counts, ground-truth alignment error distributions, or validation metrics) or description of how real-world scans were paired with models are provided. For a benchmark paper, these are load-bearing elements needed to assess reproducibility and suitability for the claimed cross-domain training and transferability analysis.

Authors: We agree that these quantitative characteristics and pairing details are essential for a benchmark paper and are not currently provided in the abstract. The manuscript mentions the public Zenodo release but lacks the specific numbers, distributions, metrics, and pairing methodology. We will revise the paper to include these elements: number of point clouds, average point counts, ground-truth alignment error distributions, validation metrics, and a description of how real-world scans were paired with models. These will be added to the abstract and a new or expanded dataset section to support reproducibility and the cross-domain claims. revision: yes

Circularity Check

0 steps flagged

No circularity: dataset introduction paper with no derivation chain

full rationale

The paper introduces the PC2Model benchmark as a publicly available hybrid dataset combining simulated point clouds and real-world scans for PC2Model registration tasks. No mathematical derivations, equations, fitted parameters, or predictions are present that could reduce to their own inputs by construction. Claims about supporting sim-to-real transferability analysis are forward-looking assertions about dataset utility rather than self-referential results derived from prior steps in the paper. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing elements. The work is self-contained as a data release and benchmark description.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a dataset and benchmark paper rather than a theoretical derivation. No free parameters, axioms, or invented entities are required to support the central claim.

pith-pipeline@v0.9.0 · 5579 in / 1052 out tokens · 41712 ms · 2026-05-10T03:35:09.964247+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

36 extracted references · 8 canonical work pages · 1 internal anchor

[1]

Applied Sciences , VOLUME =

Mora, Rocio and Martín-Jiménez, Jose Antonio and Lagüela, Susana and González-Aguilera, Diego , TITLE =. Applied Sciences , VOLUME =. 2021 , NUMBER =

2021
[2]

Surface Reconstruction From Point Clouds: A Survey and a Benchmark , year=

Huang, ZhangJin and Wen, Yuxin and Wang, ZiHao and Ren, Jinjuan and Jia, Kui , journal=. Surface Reconstruction From Point Clouds: A Survey and a Benchmark , year=
[3]

and Mawas, K

Maboudi, M. and Mawas, K. and Gerke, M. , TITLE =. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences , VOLUME =. 2025 , PAGES =

2025
[4]

and Radanović, Marko and Khoshelham, Kourosh , title =

Zhao, Yuan and Zhao, H. and Radanović, Marko and Khoshelham, Kourosh , title =. The Photogrammetric Record , year =. doi:10.1111/phor.12428 , url=

work page doi:10.1111/phor.12428
[5]

ABC: A Big CAD Model Dataset for Geometric Deep Learning , year=

Koch, Sebastian and Matveev, Albert and Jiang, Zhongshi and Williams, Francis and Artemov, Alexey and Burnaev, Evgeny and Alexa, Marc and Zorin, Denis and Panozzo, Daniele , booktitle=. ABC: A Big CAD Model Dataset for Geometric Deep Learning , year=
[6]

Pushing the Frontiers of Unconstrained Face Detection and Recog- nition: IARPA Janus Benchmark A

Wu, Zhirong and Song, Shuran and Khosla, Aditya and Yu, Fisher and Zhang, Linguang and Tang, Xiaoou and Xiao, Jianxiong , booktitle =. 3D ShapeNets: A Deep Representation for Volumetric Shapes , year =. doi:10.1109/CVPR.2015.7298801 , url=

work page doi:10.1109/cvpr.2015.7298801 2015
[7]

and Jayaraman, Pradeep Kumar and Chu, Hang and Tian, Yunsheng and Li, Yifei and Grandi, Daniele and Sanghi, Aditya and Tran, Linh and Lambourne, Joseph G

Willis, Karl D.D. and Jayaraman, Pradeep Kumar and Chu, Hang and Tian, Yunsheng and Li, Yifei and Grandi, Daniele and Sanghi, Aditya and Tran, Linh and Lambourne, Joseph G. and Solar-Lezama, Armando and Matusik, Wojciech , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2022 , pages =

2022
[8]

Thingi10K: A Dataset of 10,000 3D-Printing Models , author=
[9]

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences , volume =

The Isprs Benchmark on Indoor Modelling – Preliminary Results , author =. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences , volume =. 2020 , pages =. doi:10.5194/isprs-archives-XLIII-B5-2020-207-2020 , url=

work page doi:10.5194/isprs-archives-xliii-b5-2020-207-2020 2020
[10]

Virtual laser scanning with HELIOS++: A novel take on ray tracing-based simulation of topographic full-waveform 3D laser scanning , journal =

Lukas Winiwarter and Alberto Manuel. Virtual laser scanning with HELIOS++: A novel take on ray tracing-based simulation of topographic full-waveform 3D laser scanning , journal =. 2022 , issn =

2022
[11]

, Journal =

Smith, J. , Journal =. Close range photogrammetry for analyzing distressed trees , Volume =
[12]

Michalis and I

P. Michalis and I. Dowman , Journal =. A Generic Model for Along-Track Stereo Sensors Using Rigorous Orbit Mechanics , Volume =
[13]

Photogrammetric Computer Vision , Year =

F. Photogrammetric Computer Vision , Year =
[14]

Smith , Booktitle =

J. Smith , Booktitle =. Remote sensing to predict volcano outbursts , Volume =
[15]

and Rottensteiner, F

Maas, A. and Rottensteiner, F. and Heipke, C. , Journal =. Classification under label noise using outdated maps , Year =
[16]

K. L. Chan and K. Qin , Journal =. Biomass burning related pollution and their contributions to the air quality in Hong Kong , Year =
[17]

and Swatantran, A

Dubayah, R.O. and Swatantran, A. and Huang, W. and Duncanson, L. and Trang, H. and Johnson, K. and Dunne, J.O. and Hurtt, G.C. , Howpublished =
[18]

Multiple View Geometry in Computer Vision , publisher=

Hartley, Richard and Zisserman, Andrew , year=. Multiple View Geometry in Computer Vision , publisher=
[19]

2026 , issn =

Evaluating confidence in geometric matching between 3D point clouds and BIM models by integrating coverage, distance, and distribution metrics , journal =. 2026 , issn =. doi:https://doi.org/10.1016/j.autcon.2025.106636 , url =

work page doi:10.1016/j.autcon.2025.106636 2026
[20]

and McKay, Neil D

Besl, P.J. and McKay, Neil D. , journal=. A method for registration of 3-D shapes , year=
[21]

Go-ICP: A Globally Optimal Solution to 3D ICP Point-Set Registration , year=

Yang, Jiaolong and Li, Hongdong and Campbell, Dylan and Jia, Yunde , journal=. Go-ICP: A Globally Optimal Solution to 3D ICP Point-Set Registration , year=
[22]

and Strasser, W

Biber, P. and Strasser, W. , booktitle=. The normal distributions transform: a new approach to laser scan matching , year=
[23]

2019 IEEE/CVF International Conference on Computer Vision (ICCV) , year =

Wang, Yue and Solomon, Justin , title =. 2019 IEEE/CVF International Conference on Computer Vision (ICCV) , year =. doi:10.1109/ICCV.2019.00362 , publisher =

work page doi:10.1109/iccv.2019.00362 2019
[24]

, title =

Wang, Yue and Solomon, Justin M. , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =
[25]

Model Based Training, Detection and Pose Estimation of Texture-Less 3D Objects in Heavily Cluttered Scenes

Hinterstoisser, Stefan and Lepetit, Vincent and Ilic, Slobodan and Holzer, Stefan and Bradski, Gary and Konolige, Kurt and Navab, Nassir. Model Based Training, Detection and Pose Estimation of Texture-Less 3D Objects in Heavily Cluttered Scenes. Computer Vision -- ACCV 2012. 2013

2012
[26]

BOP: Benchmark for 6D Object Pose Estimation

Hoda n , Tom \'a s and Michel, Frank and Brachmann, Eric and Kehl, Wadim and Buch, Anders Glent and Kraft, Dirk and Drost, Bertram and Vidal, Joel and Ihrke, Stephan and Zabulis, Xenophon and Sahin, Caner and Manhardt, Fabian and Tombari, Federico and Kim, Tae-Kyun and Matas, Ji r \'i and Rother, Carsten. BOP: Benchmark for 6D Object Pose Estimation. Comp...

2018
[27]

Qian. Open3D:. CoRR , volume =. 2018 , url=. 1801.09847 , timestamp =

work page internal anchor Pith review arXiv 2018
[28]

A review on geometry and surface inspection in 3D concrete printing , journal =

Karam Mawas and Mehdi Maboudi and Markus Gerke , keywords =. A review on geometry and surface inspection in 3D concrete printing , journal =. 2026 , issn =. doi:https://doi.org/10.1016/j.cemconres.2025.108030 , url =

work page doi:10.1016/j.cemconres.2025.108030 2026
[29]

Saiti and T

E. Saiti and T. Theoharis , keywords =. Multimodal registration across 3D point clouds and CT-volumes , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.cag.2022.06.012 , url =

work page doi:10.1016/j.cag.2022.06.012 2022
[30]

Fast Global Registration

Zhou, Qian-Yi and Park, Jaesik and Koltun, Vladlen. Fast Global Registration. Computer Vision -- ECCV 2016. 2016

2016
[31]

2019 , eprint=

PCRNet: Point Cloud Registration Network using PointNet Encoding , author=. 2019 , eprint=

2019
[32]

, Howpublished =

Gago-Silva, A. , Howpublished =
[33]

, Howpublished =

Smith, J. , Howpublished =
[34]

Range Image-based LiDAR Localization for Autonomous Vehicles , booktitle=

Chen, Xieyuanli and Vizzo, Ignacio and Labe, Thomas and Behley, Jens and Stachniss, Cyrill , year=. Range Image-based LiDAR Localization for Autonomous Vehicles , booktitle=
[35]

Poisson Surface Reconstruction for LiDAR Odometry and Mapping , year=

Vizzo, Ignacio and Chen, Xieyuanli and Chebrolu, Nived and Behley, Jens and Stachniss, Cyrill , booktitle=. Poisson Surface Reconstruction for LiDAR Odometry and Mapping , year=
[36]

MICP-L: Mesh-based ICP for Robot Localization Using Hardware-Accelerated Ray Casting , booktitle=

Mock, Alexander and Wiemann, Thomas and Pütz, Sebastian and Hertzberg, Joachim , year=. MICP-L: Mesh-based ICP for Robot Localization Using Hardware-Accelerated Ray Casting , booktitle=