Breaking the Rigid Prior: Towards Articulated 3D Anomaly Detection

Bozhong Zheng; Jinye Gan; Junye Ren; Na Ni; Xiaohao Xu; Yingna Wu; Zixuan Zhang

arxiv: 2604.26868 · v1 · submitted 2026-04-29 · 💻 cs.CV

Breaking the Rigid Prior: Towards Articulated 3D Anomaly Detection

Jinye Gan , Bozhong Zheng , Xiaohao Xu , Junye Ren , Zixuan Zhang , Na Ni , Yingna Wu This is my paper

Pith reviewed 2026-05-07 11:47 UTC · model grok-4.3

classification 💻 cs.CV

keywords 3D anomaly detectionarticulated objectssigned distance fieldspose conditioningbenchmark datasetimplicit representationspoint cloud analysis

0 comments

The pith

Articulated objects break the rigid prior used in 3D anomaly detection, requiring pose-conditioned implicit fields instead.

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

The paper shows that existing 3D anomaly detection assumes rigid geometry that can be aligned to a single canonical template, but this assumption fails for objects with joints because valid pose changes create structured variations that get mistaken for defects. To address the gap, the authors release ArtiAD, a benchmark of over 15,000 point clouds from 39 categories that includes dense joint-angle changes, six types of structural anomalies, and explicit part-motion labels. They introduce SPA-SDF, which learns a continuous pose-conditioned signed distance field that separates an articulation-independent structural prior from a Fourier-encoded joint embedding. At inference the method recovers the current joint state by minimizing reconstruction energy and then identifies anomalies as point-wise deviations from the recovered manifold, reaching 0.884 AUROC on seen configurations and 0.874 on unseen ones.

Core claim

Existing 3D anomaly detection methods assume rigid objects whose geometry can be canonicalized through registration or alignment. This prior does not hold for articulated objects, where valid pose changes induce structured geometric variations that cannot be collapsed to a single template. The authors therefore introduce the ArtiAD benchmark of 15,229 point clouds across 39 categories together with dense joint-angle annotations, part-level motion labels, and a seen/unseen articulation split. Their SPA-SDF baseline replaces the rigid prior with a continuous pose-conditioned implicit field factorized into an articulation-independent structural prior and a Fourier-encoded joint embedding; at 0.

What carries the argument

SPA-SDF, a pose-conditioned signed distance field that factorizes geometry into an articulation-independent structural prior and a Fourier-encoded joint embedding, with inference-time pose recovery performed by minimizing reconstruction energy.

If this is right

Pose-induced deformations are separated from structural defects once the joint state is recovered by energy minimization.
Anomalies are detected as point-wise deviations from the learned pose-specific manifold rather than from a single rigid template.
Performance holds for both interpolation within seen joint ranges and extrapolation to novel joint configurations.
The explicit joint and part-motion labels in ArtiAD enable direct measurement of how well pose and structure are disentangled.

Where Pith is reading between the lines

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

The same energy-minimization approach for recovering articulation could be tested on real sensor data from robotic arms or folding mechanisms where ground-truth joint angles are available.
Extending the factorization to include velocity or multi-frame consistency might improve robustness when single-frame point clouds are noisy or incomplete.
The benchmark split between seen and unseen configurations provides a direct testbed for whether implicit fields generalize better than explicit part-assembly models under novel poses.

Load-bearing premise

The articulation state can be recovered at inference by minimizing reconstruction energy on the learned pose-conditioned field even when structural anomalies are present in the input.

What would settle it

On the ArtiAD unseen-configuration split, if object-level AUROC for SPA-SDF falls to or below the level of rigid registration baselines, the advantage of pose conditioning would be falsified.

Figures

Figures reproduced from arXiv: 2604.26868 by Bozhong Zheng, Jinye Gan, Junye Ren, Na Ni, Xiaohao Xu, Yingna Wu, Zixuan Zhang.

**Figure 1.** Figure 1: Overview of ArtiAD, a large-scale benchmark for articulated 3D anomaly detection. The dataset covers diverse view at source ↗

**Figure 2.** Figure 2: Data generation pipelines of ArtiAD. (a) Normal view at source ↗

**Figure 3.** Figure 3: Overview of Shape–Pose-Aware Signed Distance Field (SPA-SDF) framework. Given an input point cloud of an view at source ↗

**Figure 4.** Figure 4: Qualitative comparison of anomaly localization view at source ↗

**Figure 5.** Figure 5: Qualitative comparison of component effectiveness view at source ↗

read the original abstract

Existing 3D anomaly detection methods are built on a rigid prior: normal geometry is pose-invariant and can be canonicalized through registration or alignment. This prior does not hold for articulated objects with hinge or sliding joints, where valid pose changes induce structured geometric variations that cannot be collapsed to a single canonical template, causing pose-induced deformations to be misidentified as anomalies while true structural defects are obscured. No existing benchmark addresses this challenge. We introduce ArtiAD, the first large-scale benchmark for articulated 3D anomaly detection, comprising 15,229 point clouds across 39 object categories with dense joint-angle variations and six structural anomaly types. Each sample is annotated with its joint configuration and part-level motion labels, enabling explicit disentanglement of pose-induced geometry from structural defects. ArtiAD also provides a seen/unseen articulation split to evaluate both interpolation and extrapolation to novel joint configurations. We propose Shape-Pose-Aware Signed Distance Field (SPA-SDF), a baseline that replaces the rigid prior with a continuous pose-conditioned implicit field, factorized into an articulation-independent structural prior and a Fourier-encoded joint embedding. At inference, the articulation state is recovered by minimizing reconstruction energy, and anomalies are identified as point-wise deviations from the learned manifold. SPA-SDF achieves 0.884 object-level AUROC on seen configurations and 0.874 on unseen configurations, substantially outperforming all rigid-based baselines. Our code and benchmark will be publicly released to facilitate future research.

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

ArtiAD gives the field a first proper benchmark for articulated 3D anomaly detection and SPA-SDF shows that explicit pose modeling beats rigid baselines, but the energy-minimization step for pose recovery is untested when anomalies are present.

read the letter

The punchline is that this work fills a real gap by creating the first benchmark for 3D anomaly detection on articulated objects and shows that modeling pose explicitly helps, but the proposed inference method has an untested weak point. The paper introduces ArtiAD with over 15,000 point clouds from 39 categories, dense joint variations, six anomaly types, and seen/unseen articulation splits. This is genuinely new; no prior benchmark handled articulated cases this way. The SPA-SDF baseline factorizes the SDF into an articulation-independent structural prior plus a Fourier joint embedding. At test time it recovers the pose by energy minimization and scores deviations from the manifold. The reported object-level AUROC of 0.884 on seen and 0.874 on unseen beats the rigid baselines, which is a solid result on the surface. What works is the explicit separation of pose-induced variation from structural defects. The annotations for joint angles and part motion make it possible to evaluate whether the model actually disentangles the two. Releasing the code and benchmark is helpful for the community. The main soft spot is the pose recovery step. Minimizing reconstruction energy to find the joint angles assumes the learned field remains reliable even when the input has structural anomalies. If an anomaly creates a spurious low-energy pose or shifts the minimum, the recovered articulation will be wrong and the anomaly scores will be off. The paper gives no analysis of optimization basins, sensitivity to the six anomaly types, or failure cases. The close seen-unseen numbers are nice but do not address this because both could rely on correct pose recovery during evaluation. I would want to see ablation on whether the energy minimization succeeds on anomalous samples. This is for people building inspection systems for robots or articulated products. A reader working on implicit representations or anomaly detection will find the benchmark useful. The work shows clear thinking about the rigid prior problem and honest engagement with the literature on SDFs and pose conditioning. It deserves a serious referee. The benchmark alone makes it worth reviewing, and the method is a reasonable starting point even if the robustness needs more evidence. Recommendation: Send it out for review rather than desk reject, but ask reviewers to focus on verifying the inference reliability.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces ArtiAD, the first large-scale benchmark for 3D anomaly detection on articulated objects, comprising 15,229 point clouds across 39 categories with dense joint-angle variations, six structural anomaly types, and seen/unseen articulation splits. It proposes SPA-SDF, a baseline method using a continuous pose-conditioned implicit signed distance field factorized into an articulation-independent structural prior and a Fourier-encoded joint embedding. At inference, articulation state is recovered by minimizing reconstruction energy over the learned field, and anomalies are identified via point-wise deviations from the manifold. SPA-SDF reports object-level AUROC of 0.884 on seen configurations and 0.874 on unseen configurations, substantially outperforming rigid-based baselines.

Significance. If the results hold, the work is significant for establishing the first dedicated benchmark and evaluation protocol for articulated 3D anomaly detection, directly addressing the failure of rigid priors on objects with hinge or sliding joints. The seen/unseen split and part-level motion labels enable clear assessment of interpolation versus extrapolation. Public release of the benchmark and code is a concrete strength that will enable reproducible follow-up research.

major comments (2)

[Abstract and §3] Abstract and §3 (method): The central inference procedure recovers articulation state by minimizing reconstruction energy on the pose-conditioned SDF. No analysis is provided of optimization robustness, initialization strategy, basin of attraction, or failure cases when the six anomaly types distort the input point cloud (e.g., by creating spurious low-energy minima). This is load-bearing for the reported AUROC gap between seen (0.884) and unseen (0.874) configurations, as incorrect pose recovery would conflate pose mismatch with structural defects.
[Abstract and §4] Abstract and §4 (experiments): The outperformance claim over rigid baselines is stated with concrete AUROC numbers, but no details are given on baseline implementations (e.g., how registration or canonicalization was adapted to articulated data), anomaly labeling protocol, or statistical significance tests. Without these, the 0.01 AUROC difference between seen and unseen splits cannot be confidently attributed to the method rather than benchmark construction choices.

minor comments (2)

[Abstract] Abstract: The total sample count (15,229) and category count (39) should be cross-checked against the exact splits and tables in §4 for consistency.
[§3] Notation: The factorization into 'articulation-independent structural prior' and 'Fourier-encoded joint embedding' is described at a high level; a short equation or diagram in §3 would clarify the exact conditioning mechanism.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment of the significance of ArtiAD and SPA-SDF, and for the constructive major comments. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (method): The central inference procedure recovers articulation state by minimizing reconstruction energy on the pose-conditioned SDF. No analysis is provided of optimization robustness, initialization strategy, basin of attraction, or failure cases when the six anomaly types distort the input point cloud (e.g., by creating spurious low-energy minima). This is load-bearing for the reported AUROC gap between seen (0.884) and unseen (0.874) configurations, as incorrect pose recovery would conflate pose mismatch with structural defects.

Authors: We agree that the current manuscript lacks a dedicated analysis of the optimization procedure. In the revised version we will add a new subsection in §3 describing the initialization (mean joint angles from the training distribution), the use of multiple random restarts (typically 5) to mitigate local minima, and quantitative convergence statistics on both normal and anomalous inputs. We will also report per-anomaly-type success rates and discuss how optimization failures on specific structural defects (e.g., missing parts) are handled by the downstream point-wise deviation scoring. These additions will clarify that the modest 0.01 AUROC gap between seen and unseen splits is not an artifact of unreliable pose recovery. revision: yes
Referee: [Abstract and §4] Abstract and §4 (experiments): The outperformance claim over rigid baselines is stated with concrete AUROC numbers, but no details are given on baseline implementations (e.g., how registration or canonicalization was adapted to articulated data), anomaly labeling protocol, or statistical significance tests. Without these, the 0.01 AUROC difference between seen and unseen splits cannot be confidently attributed to the method rather than benchmark construction choices.

Authors: We acknowledge that the experimental section is insufficiently detailed. In the revision we will expand §4 with: (i) explicit descriptions of how each rigid baseline was adapted (part-aware ICP for registration where feasible, and canonicalization performed per rigid component rather than globally); (ii) the precise procedural generation protocol used to create the six anomaly types; and (iii) bootstrap-derived 95% confidence intervals on all reported AUROCs together with a paired significance test between seen and unseen splits. While the primary performance advantage of SPA-SDF is the large margin over rigid methods, these additions will allow readers to evaluate the small seen/unseen gap with appropriate statistical context. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or evaluation chain

full rationale

The paper introduces an independent benchmark (ArtiAD) with held-out seen/unseen articulation splits and defines SPA-SDF as a new pose-conditioned implicit representation whose training objective and inference procedure (energy minimization for pose recovery followed by point-wise deviation scoring) are stated directly without reducing any reported AUROC to a fitted parameter, self-citation, or input by construction. No uniqueness theorem, ansatz smuggling, or renaming of prior results is invoked as load-bearing; the central performance numbers are obtained from standard evaluation on external test data.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are detailed beyond standard use of signed distance fields and Fourier feature encodings for conditioning. The factorization into structural prior and joint embedding is presented as a modeling choice without further justification or independent evidence.

pith-pipeline@v0.9.0 · 5581 in / 1344 out tokens · 89658 ms · 2026-05-07T11:47:06.352866+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

[1]

Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. 2022. The MVTec 3D-AD Dataset for Unsupervised 3D Anomaly Detection and Localization. In 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, Vol. 5: VISAPP. Scitepress, Setúbal

work page 2022
[2]

Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. 2022. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. InProceedings of the Asian Conference on Computer Vision. 3586–3602

work page 2022
[3]

Wei Cao, Chang Luo, Biao Zhang, Matthias Nießner, and Jiapeng Tang. 2024. Mo- tion2VecSets: 4D Latent Vector Set Diffusion for Non-rigid Shape Reconstruction and Tracking. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. https://vveicao.github.io/projects/Motion2VecSets/

work page 2024
[4]

Yunkang Cao, Xiaohao Xu, and Weiming Shen. 2024. Complementary pseudo multimodal feature for point cloud anomaly detection.Pattern Recognition156 (Dec. 2024), 110761. doi:10.1016/j.patcog.2024.110761

work page doi:10.1016/j.patcog.2024.110761 2024
[5]

Xintao Chen, Xiaohao Xu, Bozhong Zheng, Yun Liu, and Yingna Wu. 2026. Un- supervised Multi-View Visual Anomaly Detection via Progressive Homography- Guided Alignment. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 3065–3073

work page 2026
[6]

Jiayi Cheng, Can Gao, Jie Zhou, Jiajun Wen, Tao Dai, and Jinbao Wang. 2025. MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence (IJCAI-2025). International Joint Conferences on Artificial Intelligence Organization, 837–8...

work page doi:10.24963/ijcai.2025/94 2025
[7]

Yuqi Cheng, Yihan Sun, Hui Zhang, Weiming Shen, and Yunkang Cao. 2026. Towards high-resolution 3d anomaly detection: A scalable dataset and real-time framework for subtle industrial defects.AAAI Conference on Artificial Intelligence (2026)

work page 2026
[8]

Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. 2021. PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization. InPattern Recognition. ICPR International Workshops and Challenges, Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, an...

work page 2021
[9]

Yao Gu, Xiaohao Xu, and Yingna Wu. 2026. Multi-turn Physics-informed Vision- language Model for Physics-grounded Anomaly Detection.IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)(2026)

work page 2026
[10]

Eliahu Horwitz and Yedid Hoshen. 2022. An Empirical Investigation of 3D Anomaly Detection and Segmentation.arXiv preprint arXiv:2203.05550(2022)

work page arXiv 2022
[11]

Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. 2022. Registration based few-shot anomaly detection. InEuro- pean Conference on Computer Vision. Springer, 303–319

work page 2022
[12]

Jiahui Lei and Kostas Daniilidis. 2022. CaDeX: Learning Canonical Deformation Coordinate Space for Dynamic Surface Representation via Neural Homeomor- phism. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. https://cis.upenn.edu/~leijh/projects/cadex

work page 2022
[13]

Jiahui Lei, Congyue Deng, William B Shen, Leonidas J Guibas, and Kostas Dani- ilidis. 2023. Nap: Neural 3d articulated object prior.Advances in Neural Informa- tion Processing Systems36 (2023), 31878–31894

work page 2023
[14]

Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. 2024. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22207– 22216

work page 2024
[15]

Wenqiao Li, Bozhong Zheng, Xiaohao Xu, Jinye Gan, Fading Lu, Xiang Li, Na Ni, Zheng Tian, Xiaonan Huang, Shenghua Gao, et al. 2025. Multi-sensor object anomaly detection: Unifying appearance, geometry, and internal properties. In Proceedings of the computer vision and pattern recognition conference. 9984–9993

work page 2025
[16]

Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. 2020. Category-level articulated object pose estimation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3706–3715

work page 2020
[17]

Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. 2023. Paris: Part-level re- construction and motion analysis for articulated objects. InProceedings of the IEEE/CVF International Conference on Computer Vision. 352–363

work page 2023
[18]

Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. 2023. Real3d-ad: A dataset of point cloud anomaly detection.Advances in Neural Information Processing Systems36 (2023), 30402– 30415

work page 2023
[19]

Srinivasan, Matthew Tancik, Jonathan T

Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. InECCV

work page 2020
[20]

Yuille, Nuno Vasconcelos, and Xiaolong Wang

Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan L. Yuille, Nuno Vasconcelos, and Xiaolong Wang. 2021. A-SDF: Learning Disentangled Signed Distance Functions for Articulated Shape Representation. (2021), 12981–12991

work page 2021
[21]

Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. 2019. Deepsdf: Learning continuous signed distance functions for shape representation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 165–174

work page 2019
[22]

Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer

work page
[23]

D-NeRF: Neural Radiance Fields for Dynamic Scenes.arXiv preprint arXiv:2011.13961(2020)

work page arXiv 2011
[24]

Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. 2022. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14318–14328

work page 2022
[25]

Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. 2023. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8032–8041

work page 2023
[26]

Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. 2020. Sapien: A simulated part-based interactive environment. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11097–11107

work page 2020
[27]

Jianan Ye, Weiguang Zhao, Xi Yang, Guangliang Cheng, and Kaizhu Huang. 2025. Po3ad: Predicting point offsets toward better 3d point cloud anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference. 1353–1362

work page 2025
[28]

Bozhong Zheng, Jinye Gan, Xiaohao Xu, Xintao Chen, Wenqiao Li, Xiaonan Huang, Na Ni, and Yingna Wu. 2025. Bridging 3d anomaly localization and repair via high-quality continuous geometric representation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 27063–27072. Preprint, April, 2026 Gan et al

work page 2025
[29]

Qihang Zhou, Jiangtao Yan, Shibo He, Wenchao Meng, and Jiming Chen. 2024. PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems. https://openreview.net/forum?id=02CIZ8qeDc

work page 2024
[30]

Zheyuan Zhou, Le Wang, Naiyu Fang, Zili Wang, Lemiao Qiu, and Shuyou Zhang

work page
[31]

In European Conference on Computer Vision

R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection. In European Conference on Computer Vision. Springer, 91–107. A ArtiAD Dataset Details A.1 Articulation Space Coverage ArtiAD explicitly models articulation as a continuous variable. For hinge-based objects, the articulation parameter 𝜃 is sampled from a bounded angular range (e.g., [0◦, 120◦ ]...

work page arXiv 2026

[1] [1]

Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. 2022. The MVTec 3D-AD Dataset for Unsupervised 3D Anomaly Detection and Localization. In 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, Vol. 5: VISAPP. Scitepress, Setúbal

work page 2022

[2] [2]

Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. 2022. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. InProceedings of the Asian Conference on Computer Vision. 3586–3602

work page 2022

[3] [3]

Wei Cao, Chang Luo, Biao Zhang, Matthias Nießner, and Jiapeng Tang. 2024. Mo- tion2VecSets: 4D Latent Vector Set Diffusion for Non-rigid Shape Reconstruction and Tracking. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. https://vveicao.github.io/projects/Motion2VecSets/

work page 2024

[4] [4]

Yunkang Cao, Xiaohao Xu, and Weiming Shen. 2024. Complementary pseudo multimodal feature for point cloud anomaly detection.Pattern Recognition156 (Dec. 2024), 110761. doi:10.1016/j.patcog.2024.110761

work page doi:10.1016/j.patcog.2024.110761 2024

[5] [5]

Xintao Chen, Xiaohao Xu, Bozhong Zheng, Yun Liu, and Yingna Wu. 2026. Un- supervised Multi-View Visual Anomaly Detection via Progressive Homography- Guided Alignment. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 3065–3073

work page 2026

[6] [6]

Jiayi Cheng, Can Gao, Jie Zhou, Jiajun Wen, Tao Dai, and Jinbao Wang. 2025. MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence (IJCAI-2025). International Joint Conferences on Artificial Intelligence Organization, 837–8...

work page doi:10.24963/ijcai.2025/94 2025

[7] [7]

Yuqi Cheng, Yihan Sun, Hui Zhang, Weiming Shen, and Yunkang Cao. 2026. Towards high-resolution 3d anomaly detection: A scalable dataset and real-time framework for subtle industrial defects.AAAI Conference on Artificial Intelligence (2026)

work page 2026

[8] [8]

Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. 2021. PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection and Localization. InPattern Recognition. ICPR International Workshops and Challenges, Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, an...

work page 2021

[9] [9]

Yao Gu, Xiaohao Xu, and Yingna Wu. 2026. Multi-turn Physics-informed Vision- language Model for Physics-grounded Anomaly Detection.IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP)(2026)

work page 2026

[10] [10]

Eliahu Horwitz and Yedid Hoshen. 2022. An Empirical Investigation of 3D Anomaly Detection and Segmentation.arXiv preprint arXiv:2203.05550(2022)

work page arXiv 2022

[11] [11]

Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. 2022. Registration based few-shot anomaly detection. InEuro- pean Conference on Computer Vision. Springer, 303–319

work page 2022

[12] [12]

Jiahui Lei and Kostas Daniilidis. 2022. CaDeX: Learning Canonical Deformation Coordinate Space for Dynamic Surface Representation via Neural Homeomor- phism. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. https://cis.upenn.edu/~leijh/projects/cadex

work page 2022

[13] [13]

Jiahui Lei, Congyue Deng, William B Shen, Leonidas J Guibas, and Kostas Dani- ilidis. 2023. Nap: Neural 3d articulated object prior.Advances in Neural Informa- tion Processing Systems36 (2023), 31878–31894

work page 2023

[14] [14]

Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. 2024. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22207– 22216

work page 2024

[15] [15]

Wenqiao Li, Bozhong Zheng, Xiaohao Xu, Jinye Gan, Fading Lu, Xiang Li, Na Ni, Zheng Tian, Xiaonan Huang, Shenghua Gao, et al. 2025. Multi-sensor object anomaly detection: Unifying appearance, geometry, and internal properties. In Proceedings of the computer vision and pattern recognition conference. 9984–9993

work page 2025

[16] [16]

Xiaolong Li, He Wang, Li Yi, Leonidas J Guibas, A Lynn Abbott, and Shuran Song. 2020. Category-level articulated object pose estimation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3706–3715

work page 2020

[17] [17]

Jiayi Liu, Ali Mahdavi-Amiri, and Manolis Savva. 2023. Paris: Part-level re- construction and motion analysis for articulated objects. InProceedings of the IEEE/CVF International Conference on Computer Vision. 352–363

work page 2023

[18] [18]

Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. 2023. Real3d-ad: A dataset of point cloud anomaly detection.Advances in Neural Information Processing Systems36 (2023), 30402– 30415

work page 2023

[19] [19]

Srinivasan, Matthew Tancik, Jonathan T

Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. InECCV

work page 2020

[20] [20]

Yuille, Nuno Vasconcelos, and Xiaolong Wang

Jiteng Mu, Weichao Qiu, Adam Kortylewski, Alan L. Yuille, Nuno Vasconcelos, and Xiaolong Wang. 2021. A-SDF: Learning Disentangled Signed Distance Functions for Articulated Shape Representation. (2021), 12981–12991

work page 2021

[21] [21]

Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. 2019. Deepsdf: Learning continuous signed distance functions for shape representation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 165–174

work page 2019

[22] [22]

Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer

work page

[23] [23]

D-NeRF: Neural Radiance Fields for Dynamic Scenes.arXiv preprint arXiv:2011.13961(2020)

work page arXiv 2011

[24] [24]

Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. 2022. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14318–14328

work page 2022

[25] [25]

Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. 2023. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8032–8041

work page 2023

[26] [26]

Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. 2020. Sapien: A simulated part-based interactive environment. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11097–11107

work page 2020

[27] [27]

Jianan Ye, Weiguang Zhao, Xi Yang, Guangliang Cheng, and Kaizhu Huang. 2025. Po3ad: Predicting point offsets toward better 3d point cloud anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference. 1353–1362

work page 2025

[28] [28]

Bozhong Zheng, Jinye Gan, Xiaohao Xu, Xintao Chen, Wenqiao Li, Xiaonan Huang, Na Ni, and Yingna Wu. 2025. Bridging 3d anomaly localization and repair via high-quality continuous geometric representation. InProceedings of the IEEE/CVF International Conference on Computer Vision. 27063–27072. Preprint, April, 2026 Gan et al

work page 2025

[29] [29]

Qihang Zhou, Jiangtao Yan, Shibo He, Wenchao Meng, and Jiming Chen. 2024. PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems. https://openreview.net/forum?id=02CIZ8qeDc

work page 2024

[30] [30]

Zheyuan Zhou, Le Wang, Naiyu Fang, Zili Wang, Lemiao Qiu, and Shuyou Zhang

work page

[31] [31]

In European Conference on Computer Vision

R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection. In European Conference on Computer Vision. Springer, 91–107. A ArtiAD Dataset Details A.1 Articulation Space Coverage ArtiAD explicitly models articulation as a continuous variable. For hinge-based objects, the articulation parameter 𝜃 is sampled from a bounded angular range (e.g., [0◦, 120◦ ]...

work page arXiv 2026