Pith. sign in

REVIEW 5 major objections 6 minor 62 references

Examining the Source of Defects from a Mechanical Perspective for 3D Anomaly Detection

T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read This paper argues that every 3D surface defect can be modeled as an imbalance of internal and external mechanical forces, and that a network trained to predict the corrective force at each point yields anomaly scores that beat…

desk verdict A solid, incremental 3D anomaly detection paper with real efficiency gains and a useful dataset, but the pseudo-anomaly generator has a likely center-zero bug and the 'mechanical' theory is motivating metaphor rather than physics. read the letter →

arxiv 2505.05901 v2 pith:ZLG4UQAP submitted 2025-05-09 cs.CV cs.AI

classification cs.CVcs.AI
keywords 3Danomalydetectionpointcloudmechanicalforcescorrectiveforcepredictionpseudo-anomalygenerationindustrialqualitycontrolAnomaly-IntraVariancedatasetMvTec3D-AD
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Most 3D anomaly detectors describe what a defect looks like, comparing test geometry against memorized normal geometry. This paper claims that the right question is why the defect exists: every anomaly is the trace of an unpredictable defective force, either internal (for instance expanding gas) or external (for instance a collision), acting on the product surface. It proposes MC4AD, which synthesizes pseudo-anomalies along surface normals, trains a lightweight U-Net to predict the corrective force vector that would push each point back to its normal position, and scores each point by the magnitude of that predicted force. The authors report that this mechanical inversion obtains nine state-of-the-art detection and segmentation results across five established benchmarks plus their new Anomaly-IntraVariance dataset, using 14M parameters and the fastest inference among compared methods. If the claims hold, the source of a defect, not just its shape, is enough to find it.

What carries the argument

The load-bearing object is a pair of theoretical definitions: a defect force $F_D = \int_{S^+} f_E \, dS^+ + \int_{S^-} f_I \, dS^-$ that deforms a normal manifold $M$ into $M'$, and its opposite corrective force $F_C = -F_D$ that restores it. The whole pipeline is built to predict $F_C$: DA-Gen synthesizes the pseudo-defect deformation, CFP-Net (a differentiable MinkUNet-style U-Net with MC-Skip complementary feature mixing) emits the six-channel per-point corrective force, and the combined loss $L_{\text{comb}}=L_{\text{dist}}+L_{\text{dir}}+L_{\text{sym}}$ constrains it to match the synthesized deformation and to keep the internal and external components anti-symmetric. The Hierarchical Quality Control strategy adds a pruned 60%-smaller network that pre-screens obviously normal samples before the full network runs.

What would settle it

Take a defect class absent from DA-Gen's family, such as scratches that displace points tangentially rather than normally, or surface cracks visible only in texture while depth stays flat, and measure per-point AUROC on that class alone. If the corrective-force norm separates such points no better than chance while matching the paper's numbers on bulge and dent classes, the mechanical premise is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that anomaly detection can be reframed as force-equilibrium restoration. In the Mechanics Complementary Model, a normal point is in equilibrium: both defect force $F_D$ and corrective force $F_C$ are zero. A defective point is subject to a nonzero resultant force $F_D = F_E + F_I$ (external and internal components), which deforms the surface by a nonlinear map $\phi(F, M) = \nabla M$; the corrective force is exactly its opposite, $F_C = -F_D$. Training therefore does not rely on normal features alone: DA-Gen deforms training patches by perturbing surface normals with controlled magnitude $\gamma \in [0.06, 0.12]$, normal-dominance $\lambda \in [0.95, 1]$, and stretching $\sigma \in [0, 0.08]$, producing realistic pseudo-defects; CFP-Net, a differentiable sparse U-Net with complementary MC-Skip connections that mix internal and external feature streams, predicts a six-channel corrective force (3D external plus 3D internal) per point. The combined loss enforces that the predicted force reconstructs the known pseudo-defect ($L_{\text{dist}}$, $L_{\text{dir}}$) and that internal and external corrective components are anti-symmetric in anomalous regions ($L_{\text{sym}}$), which the ablation singles out as the largest single contributor. At inference the norm $\|F_C\|$ is the point anomaly score and its maximum is the object score. Reported numbers include O-AUROC 90.9% on Anomaly-ShapeNet, 95.4% on MvTec3D-AD, 78.6% on Real3D-AD, 88.8% on Anomaly-ShapeNet-New, and 76.1%/62.5% on the two groups of the new Anomaly-IntraVariance dataset.

Load-bearing premise

Every real defect is assumed to be a small local displacement of surface points roughly along the surface normal, with magnitude between 0.06 and 0.12 and a specific spatial attenuation profile, so that for every anomaly there is a corrective force whose magnitude is a reliable score; defects outside this deformation family are outside the model's training signal.

Editorial extensions

If this is right

  • The corrective-force magnitude can replace expensive memory-bank or reconstruction comparisons entirely, since inference is a single forward pass over the point cloud.
  • Splitting the score into external and internal channels is not decorative: Figure 3 shows that fusing both channels lifts O-AUROC from about 0.855/0.843 to 0.912, which suggests the complementary force channels carry partly independent signal.
  • The symmetry loss is a real lever: removing $L_{\text{sym}}$ hurts O-AUROC by 13.4% (from 0.786 to 0.652), so the mechanical constraint is doing genuine work.
  • A pruned 60%-smaller model can serve as a first-pass filter in HQC, speeding up inference (up to 26.4 FPS on Anomaly-ShapeNet) with essentially no accuracy loss, which fits cost-limited assembly-line inspection.
  • Benchmarking on Anomaly-IntraVariance shows that intraclass variance is hard: Group 2 (four subspecies per class) drops all methods to 0.625 or below, framing a realistic target for future work.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The mechanical vocabulary is likely an interpretative scaffold: the same arithmetic, predicting a per-point residual offset and scoring by its norm, is at heart what offset-based detectors do; the distinct contribution here is the internal/external decomposition and the anti-symmetry constraint. A direct control that trains a two-channel residual model without the force framing would separate the
  • A targeted experiment the paper does not run: apply the trained CFP-Net to a physical simulation, such as finite-element stress on a scanned part, and compare predicted corrective-force directions with principal strain directions. Agreement would evidence the mechanical interpretation; disagreement would mean 'force' is only a naming convention.
  • Because DA-Gen only displaces points near the surface normal with magnitude in [0.06, 0.12], defects without geometric displacement, such as color or texture changes, subsurface voids, or material-phase changes, are outside the training distribution; a dedicated test set of such non-geometric defects would reveal whether the premise extends beyond morphology.
  • The paper itself concedes in Section 6 and Appendix H that it lacks explicit physical constraints to derive damage forces, so the force vocabulary is learned implicitly from synthesized deformations and the mechanical interpretation rests on how well the pseudo-anomaly distribution matches real defects.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes MC4AD, a 3D anomaly detection framework that reinterprets defects as the result of internal and external 'damage forces' and trains a network to predict opposing 'corrective forces' whose magnitude serves as the anomaly score. The method contains three main components: DA-Gen, which generates pseudo-anomalies by displacing points along perturbed surface normals; CFP-Net, a Minkowski U-Net with complementary skip connections that outputs a six-channel corrective force field; and a combined loss with symmetry, distance, and directional terms. The paper also introduces an HQC pruning strategy and a new Anomaly-IntraVariance dataset with intraclass variance. Experiments are reported on Anomaly-ShapeNet, Anomaly-ShapeNet-New, MvTec3D-AD, Real3D-AD, and the new dataset, claiming nine state-of-the-art results with low memory and high inference speed.

Significance. If the mechanical framing were valid, the work would offer a conceptually novel perspective on 3D anomaly detection and a practical efficiency-oriented framework. The paper includes public code, extensive comparisons across many categories, a resource analysis, a generalizable hierarchical quality-control module, and a new dataset, which are useful contributions to the community. However, the central theoretical claim is currently not supported: the 'damage force' is defined as the displacement produced by DA-Gen, the losses as written drive the network to reproduce that same displacement rather than its opposite, and the anomaly score is therefore a fitted reconstruction of the authors' own synthetic augmentation. The empirical gains are also close to noise on some benchmarks, and no error bars or out-of-family transfer tests are provided. The dataset and efficiency experiments are solid enough to be of value, but the paper's central claim needs substantial revision.

major comments (5)
  1. [Section 4.1, Eq. (5)] There is an internal inconsistency at the center of the generated anomaly. In Eq. (5), FD,i is proportional to (pi_i/max(pi_i)) * (1 - sigma_i * |pi_i|), where pi_i is the normalized projected distance to the anomaly center. At the center pi_i = 0, so the first factor is zero and the generated displacement is zero, while the text states that gamma specifies the maximum displacement magnitude at the anomaly center. This means pseudo-anomalies are ring-like with an unmoved core. Since DA-Gen is the sole training signal, a network trained on these samples may assign low scores to the center of real defects, directly undermining the use of the maximum point score as the sample-level anomaly score. The authors should either define pi with an offset so that the center receives the maximum displacement, or explicitly analyze and justify the ring-like deformation.
  2. [Section 4.3, Eq. (10)] The loss terms in Eq. (10) contradict the stated mechanics. The text and Eq. (4) require FC = -FD, i.e., the corrective force opposes the damage force. However, Ldist minimizes ||FD,i - FC,i||, which drives FC toward +FD, and Ldir contains the negative normalized dot product of FD and FC, which when minimized also aligns FC with +FD. Similarly, Lsym contains the negative dot product of F'I and F'E, which aligns the two corrective components rather than making them opposing, contrary to the claim in Section 4.3 that the symmetry loss enforces opposing directions. As written, the combined loss trains the network to reconstruct the DA-Gen displacement field. The authors must correct the signs in Eq. (10) or substantially reframe the method as pure pseudo-anomaly reconstruction without the mechanical corrective-force narrative.
  3. [Sections 4.1 and 4.3] The core 'prediction' is circular with respect to the anomaly score. The damage force FD,i in Eq. (5) is defined as the displacement applied by DA-Gen, and the network is trained with a loss that minimizes the difference between FD,i and the predicted corrective force FC,i. At inference, the norm of FC is used as the anomaly score. Thus the score measures how well the network reconstructs the authors' synthetic augmentation field; no independent physical constraint enters. The ablations in Table 9 only vary parameters inside the DA-Gen family (G, lambda, sigma, gamma). To support the claim of generalization to real defects, the paper needs experiments where the training distribution and the test defect distribution are explicitly different in kind (e.g., non-normal deformations, material-level changes, larger displacement magnitudes, or real scanned defects), and an analysis of whether real defect displacements lie in the support of Eq. (5).
  4. [Section 5.2 and Tables 16-18] The experimental evidence for state-of-the-art claims is weakened by missing variance estimates and close margins. On MvTec3D-AD the O-AUROC is 0.954 versus 0.951 for the second-best method, and on Real3D-AD the P-AUROC is 0.837 versus 0.836; with no error bars or multiple seeds it is unclear whether these differences are significant. In addition, the text in Section 5.2 states that MC4AD outperforms 'M3DM of 94.5% and 90.6%' on MvTec3D-AD, but Tables 16 and 17 report M3DM average values of 0.874 O-AUROC and 0.906 P-AUROC, so the cited 94.5% is inconsistent with the tables. The same paragraph says the comparison spans 12 categories while Table 16 lists 10 categories. These inconsistencies and missing statistical evidence should be resolved before the 'nine state-of-the-art performances' claim can be accepted.
  5. [Appendix B.3, Proposition A1 and abstract] The theoretical support is overstated. Proposition A1 shows that the network mapping is differentiable, which is a generic property of convolutional networks with smooth activations and does not establish that the norm of a predicted corrective force is a valid anomaly score. Lemma A2 is cited from a preprint titled 'Reconstruct Anything Model' and is not a formal theorem about U-Nets solving arbitrary inverse problems in anomaly detection. The abstract states that MC4AD 'has been proven effective through theory and experimentation,' but the provided theory neither proves effectiveness nor connects the mechanical force model to detection guarantees. The authors should either provide a genuine theoretical statement linking the force-field objective to anomaly separability, or remove the 'proven by theory' claim.
minor comments (6)
  1. [Section 1 heading] The heading contains a typo: 'Introductuon' should be 'Introduction'.
  2. [Section 4, first paragraph] The model name is inconsistently written as 'CTF-Net' in the text and 'CFP-Net' in the figure and elsewhere; please unify.
  3. [Section 5.2] The text says MvTec3D-AD has 12 categories, but Table 16 lists 10 categories; please correct the count in the text.
  4. [Appendix G] The dataset name is written as 'Mvtec-AD' in one place; use the consistent form 'MvTec3D-AD'.
  5. [Appendix H] The heading 'Limitations and Feature Work' should read 'Limitations and Future Work'.
  6. [Section 4.3, last sentence] The final sentence of Section 4.3 ('with the reconstruction effect') is grammatically incomplete and should be rephrased.

Circularity Check

1 steps flagged · score 6.0 of 10

The corrective-force anomaly score is the norm of a network trained to reproduce DA-Gen's own synthetic displacement, making the core prediction self-referential.

  1. self definitional [Sec. 4.1 Eq. (5); Sec. 4.2 "Anomaly Score during Inference"; Sec. 4.3 Eq. (10)]
    "The norm of point-level corrective resultant force FC is directly considered as an anomaly score during the inference phase. ... Lcomb = Ldist + Ldir + Lsym = ... + 1/N Σ ∥FD,i−FC,i∥ − 1/N Σ (FD,i/∥FD,i∥2+ϵ)·(FC,i/∥FC,i∥2+ϵ), (10)"

    By Eq. 10, FC is trained with the distance term ∥FD−FC∥ and directional alignment term (FD/∥FD∥)·(FC/∥FC∥), while Eq. 4 declares FC = −FD. The 'corrective force' whose norm is the inference anomaly score is therefore, by construction, the DA-Gen pseudo-defect displacement of Eq. 5. The network is a regressor from input point clouds to the synthetic displacement field used to create its own training anomalies. The rule that normal points have zero force and anomalous points nonzero force is not derived from mechanics but is a label assigned by DA-Gen. Hence the central claim—force magnitude as a defect score—reduces to the assumption that real defects lie in the DA-Gen displacement family, an assumption the ablations only vary within that same family.

full rationale

The derivation chain from defects to corrective forces is purely definitional: the paper posits that anomalies arise from a damage force FD (Def. A2), defines the corrective force as its negative (Eq. 4), then trains CFP-Net to output FC that minimizes ∥FD−FC∥ (Eq. 10). The anomaly score is ∥FC∥. Thus the score is the magnitude of the DA-Gen pseudo-defect displacement, i.e., the network is fitted to invert the paper's own synthetic augmentation. This is not a first-principles prediction of defect forces; it is a reconstruction of a predefined augmentation. The paper's own limitation statement confirms 'MC4AD lacks explicit physical constraints to derive damage forces'. The empirical gains are against external benchmarks, but the core mechanism remains self-referential. However, the paper does compare fairly against other methods, and the HQC/dataset contributions are independent, so the overall circularity score is moderate (6) rather than extreme.

Assumptions & free parameters 6 free parameters · 4 assumptions · 3 invented entities

The central claim relies on several unverified assumptions: the mechanical causality of defects, the representativeness of synthetic anomalies, and the justification of a U-Net via an informal universality claim. The DA-Gen hyperparameters are tuned on the target datasets, and the 'forces' are unvalidated invented entities. This ledger shows that the 'theory' is a narrative overlay on a supervised reconstruction network, and the empirical success depends on the fidelity of the synthetic defect distribution.

free parameters (6)
  • DA-Gen max displacement gamma = random in [0.06, 0.12]
    Controls the magnitude of the pseudo-anomaly displacement; chosen by ablation on Real3D-AD (Appendix E).
  • DA-Gen normal dominance lambda = random in [0.95, 1]
    Balances normal-direction vs random-direction displacement; tuned on Real3D-AD (Appendix E).
  • DA-Gen stretching sigma = random in [0, 0.08]
    Controls scratch width; tuned on Real3D-AD (Appendix E).
  • Number of anomaly patches G = 64
    Size of pseudo-anomaly regions; ablation on Real3D-AD shows best at 64 (Appendix E.1).
  • HQC threshold b = 0.25
    Fraction of samples classified as normal in the first stage; tuned on Real3D-AD (Appendix E.3).
  • Voxel size = 0.03
    Voxelization resolution for the sparse U-Net, taken from MinkUNet standard configuration (Appendix A.5).
assumptions (4)
  • domain assumption Anomalies arise from unpredictable defective forces from internal and external sources.
    This is an untested causal metaphor; the paper provides no physical measurement or external validation of such forces.
  • ad hoc to paper The effect of a force on a point cloud is an additive displacement, and corrective forces restore by being exactly opposite to damage forces.
    The additive map phi in Definition A4 is never concretely defined; this is a modeling choice to justify the network output, not a derived physical law.
  • domain assumption A U-Net is universally applicable for arbitrary inverse generative tasks.
    Lemma A2 is cited from an arXiv paper (Reconstruct Anything Model), not a standard theorem; the claim is informal.
  • domain assumption DA-Gen's synthetic anomalies are representative of real defects in industrial point clouds.
    The model is trained exclusively on these pseudo-anomalies; if real defects differ in structure, the anomaly score may fail. The paper's own limitation section acknowledges the lack of physical constraints.
invented entities (3)
  • Internal corrective force F'_I
    purpose: A per-point 3D vector representing the model's predicted correction from internal sources.
    This is a network output, never measured or compared against any physical internal stress or strain.
  • External corrective force F'_E
    purpose: A per-point 3D vector representing the model's predicted correction from external sources.
    Same as internal force; it is a learned vector field with no physical validation.
  • Damage force F_D
    purpose: A latent quantity representing the cause of the defect; its negative is the training target for corrective forces.
    Defined purely by the authors' synthetic displacement model (Eq. 5), not by any measurement of actual defect-causing forces.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Examining the Source of Defects from a Mechanical Perspective for 3D Anomaly Detection." pith.science (2026). https://pith.science/paper/ZLG4UQAP

@misc{pith2026250505901,
  author       = {Pith},
  title        = {Pith review of: Examining the Source of Defects from a Mechanical Perspective for 3D Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZLG4UQAP}},
  note         = {Machine review of arXiv:2505.05901}
}
read the original abstract

In this paper, we explore a novel approach to 3D anomaly detection (AD) that goes beyond merely identifying anomalies based on structural characteristics. Our primary perspective is that most anomalies arise from unpredictable defective forces originating from both internal and external sources. To address these anomalies, we seek out opposing forces that can help correct them. Therefore, we introduce the Mechanics Complementary Model-based Framework for the 3D-AD task (MC4AD), which generates internal and external corrective forces for each point. We first propose a Diverse Anomaly-Generation (DA-Gen) module designed to simulate various types of anomalies. Next, we present the Corrective Force Prediction Network (CFP-Net), which uses complementary representations for point-level analysis to simulate the different contributions from internal and external corrective forces. To ensure the corrective forces are constrained effectively, we have developed a combined loss function that includes a new symmetric loss and an overall loss. Notably, we implement a Hierarchical Quality Control (HQC) strategy based on a three-way decision process and contribute a dataset titled Anomaly-IntraVariance, which incorporates intraclass variance to evaluate our model. As a result, the proposed MC4AD has been proven effective through theory and experimentation. The experimental results demonstrate that our approach yields nine state-of-the-art performances, achieving optimal results with minimal parameters and the fastest inference speed across five existing datasets, in addition to the proposed Anomaly-IntraVariance dataset. The source is available at https://github.com/hzzzzzhappy/MC4AD

Figures

Figures reproduced from arXiv: 2505.05901 by the authors.

Figure 1
Figure 1. Overview of our MC4AD. Compared with existing methods, the proposed MC4AD demonstrates excellent performance, low memory, and more efficiency. We examine for the first time the source of anomalies from a mechanical perspective and use it as a motivation to design models. Abstract In this paper, we explore a novel approach to 3D anomaly detection (AD) that goes beyond merely identifying anomalies based on structural … view at source ↗
Figure 2
Figure 2. Overview of MC4AD. The proposed MC4AD contains three important parts: 1) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The distribution of anomaly detection scores on class “vase9”, compared between External, [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Localization results. Benefit of Combining External & Internal Corrective Forces. MC4AD achieves 90.9% O-AUROC by integrating bidirectional corrective forces [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the training and test sets on Group 1 and Group 2. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: The distribution of anomaly detection scores on different classes, compared between [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of localization. The realistic pseudo-anomalies facilitate our CFP-Net to capture the potential damaging forces of anomalous structures and generate corrective forces to restore the anomalies. We show the generated anomaly samples with real anomaly …
Figure 8
Figure 8. Figure 8: More qualitative results of localization. [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: More qualitative results of localization. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 36 canonical work pages

  1. [1]

    A survey on rgb, 3d, and multimodal approaches for unsupervised industrial image anomaly detection,

    Y . Lin, Y . Chang, X. Tong, J. Yu, A. Liotta, G. Huang, W. Song, D. Zeng, Z. Wu, Y . Wang, and W. Zhang, “A survey on rgb, 3d, and multimodal approaches for unsupervised industrial image anomaly detection,” Information Fusion, vol. 121, p. 103139, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S156625352500212X

  2. [2]

    Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network,

    W. Li, X. Xu, Y . Gu, B. Zheng, S. Gao, and Y . Wu, “Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 22 207–22 216

  3. [3]

    Real3d-ad: A dataset of point cloud anomaly detection,

    J. Liu, G. Xie, R. Chen, X. Li, J. Wang, Y . Liu, C. Wang, and F. Zheng, “Real3d-ad: A dataset of point cloud anomaly detection,” in Advances in Neural Information Processing Systems , A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 30 402–30 415. [Online]. Available: https://proceedin...

  4. [4]

    A 3D Multimodal Feature for Infrastructure Anomaly Detection

    Y . Jing, W. Lin, B. Sheil, and S. Acikgoz, “A 3D Multimodal Feature for Infrastructure Anomaly Detection,” arXiv e-prints, p. arXiv:2502.05779, Feb. 2025

  5. [5]

    Real3d-ad: A dataset of point cloud anomaly detection,

    J. Liu, G. Xie, X. Li, J. Wang, Y . Liu, C. Wang, F. Zhenget al., “Real3d-ad: A dataset of point cloud anomaly detection,” in Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023

  6. [6]

    Deep Subspace Learning for Surface Anomaly Classification Based on 3D Point Cloud Data

    X. Cao, C. Tao, and J. Du, “Deep subspace learning for surface anomaly classification based on 3d point cloud data,” 2025. [Online]. Available: https://arxiv.org/abs/2502.11669

  7. [7]

    Robust distribution alignment for industrial anomaly detection under distribution shift,

    J. Liao, X. Xu, Y . Su, R.-C. Tu, Y . Liu, D. Tao, and X. Yang, “Robust distribution alignment for industrial anomaly detection under distribution shift,” 2025. [Online]. Available: https://arxiv.org/abs/2503.14910

  8. [8]

    Back to the feature: Classical 3d features are (almost) all you need for 3d anomaly detection,

    E. Horwitz and Y . Hoshen, “Back to the feature: Classical 3d features are (almost) all you need for 3d anomaly detection,” 2022. [Online]. Available: https://arxiv.org/abs/2203.05550

Show all 62 references
  1. [9]

    Look inside for more: Internal spatial modality perception for 3d anomaly detection,

    H. Liang, G. Xie, C. Hou, B. Wang, C. Gao, and J. Wang, “Look inside for more: Internal spatial modality perception for 3d anomaly detection,” 2025. [Online]. Available: https://arxiv.org/abs/2412.13461

  2. [10]

    Towards high-resolution 3d anomaly detection via group-level feature contrastive learning,

    H. Zhu, G. Xie, C. Hou, T. Dai, C. Gao, J. Wang, and L. Shen, “Towards high-resolution 3d anomaly detection via group-level feature contrastive learning,” in Proceedings of the 32nd ACM International Conference on Multimedia, ser. MM ’24. ACM, Oct. 2024, p. 4680–4689. [Online]...

  3. [11]

    R3d-ad: Reconstruction via diffusion for 3d anomaly detection,

    Z. Zhou, L. Wang, N. Fang, Z. Wang, L. Qiu, and S. Zhang, “R3d-ad: Reconstruction via diffusion for 3d anomaly detection,” in European Conference on Computer Vision (ECCV), 2024

  4. [12]

    Po3ad: Predicting point offsets toward better 3d point cloud anomaly detection,

    J. Ye, W. Zhao, X. Yang, G. Cheng, and K. Huang, “Po3ad: Predicting point offsets toward better 3d point cloud anomaly detection,” 2024. [Online]. Available: https://arxiv.org/abs/2412.12617

  5. [13]

    Deep industrial image anomaly detection: A survey,

    J. Liu, G. Xie, J. Wang, S. Li, C. Wang, F. Zheng, and Y . Jin, “Deep industrial image anomaly detection: A survey,” Machine Intelligence Research, vol. 21, no. 1, p. 104–135, Jan. 2024. [Online]. Available: http://dx.doi.org/10.1007/s11633-023-1459-z

  6. [14]

    Deep learning for unsupervised anomaly localization in industrial images: A survey,

    X. Tao, X. Gong, X. Zhang, S. Yan, and C. Adak, “Deep learning for unsupervised anomaly localization in industrial images: A survey,”IEEE Transactions on Instrumentation and Mea- surement, vol. 71, pp. 1–21, 2022

  7. [15]

    Im-iad: Industrial image anomaly detection benchmark in manufacturing,

    G. Xie, J. Wang, J. Liu, J. Lyu, Y . Liu, C. Wang, F. Zheng, and Y . Jin, “Im-iad: Industrial image anomaly detection benchmark in manufacturing,” IEEE Transactions on Cybernetics, vol. 54, no. 5, pp. 2720–2733, 2024

  8. [16]

    Anoseg: Anomaly segmentation network using self-supervised learning,

    J. Song, K. Kong, Y .-I. Park, S.-G. Kim, and S.-J. Kang, “Anoseg: Anomaly segmentation network using self-supervised learning,” 2021. [Online]. Available: https: //arxiv.org/abs/2110.03396

  9. [17]

    Omni-frequency channel-selection representations for unsupervised anomaly detection,

    Y . Liang, J. Zhang, S. Zhao, R. Wu, Y . Liu, and S. Pan, “Omni-frequency channel-selection representations for unsupervised anomaly detection,” arXiv preprint arXiv:2203.00259, 2022. 24

  10. [18]

    Unsupervised anomaly detection for surface defects with dual-siamese network,

    X. Tao, D. Zhang, W. Ma, Z. Hou, Z. Lu, and C. Adak, “Unsupervised anomaly detection for surface defects with dual-siamese network,” IEEE Transactions on Industrial Informatics, vol. 18, no. 11, pp. 7707–7717, 2022

  11. [19]

    Simple and effective frequency-aware image restoration for industrial visual anomaly detection,

    T. Liu, B. Li, X. Du, B. Jiang, L. Geng, F. Wang, and Z. Zhao, “Simple and effective frequency-aware image restoration for industrial visual anomaly detection,” Advanced Engineering Informatics , vol. 64, p. 103064, Mar. 2025. [Online]. Available: http://dx.doi.org/10.1016/j.a...

  12. [20]

    Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise,

    J. Wyatt, A. Leach, S. M. Schmon, and C. G. Willcocks, “Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022, pp. 649–655

  13. [21]

    A diffusion- based framework for multi-class anomaly detection,

    H. He, J. Zhang, H. Chen, X. Chen, Z. Li, X. Chen, Y . Wang, C. Wang, and L. Xie, “A diffusion- based framework for multi-class anomaly detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 8, 2024, pp. 8472–8480

  14. [22]

    Towards total recall in industrial anomaly detection,

    K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox, and P. Gehler, “Towards total recall in industrial anomaly detection,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 14 298–14 308

  15. [23]

    Sub-image anomaly detection with deep pyramid correspondences,

    N. Cohen and Y . Hoshen, “Sub-image anomaly detection with deep pyramid correspondences,”

  16. [24]

    Efficientad: Accurate visual anomaly detection at millisecond-level latencies,

    K. Batzner, L. Heckler, and R. König, “Efficientad: Accurate visual anomaly detection at millisecond-level latencies,” 2024. [Online]. Available: https://arxiv.org/abs/2303.14535

  17. [25]

    Uninformed students: Student- teacher anomaly detection with discriminative latent embeddings,

    M. Bergmann, Pauland Fauser, D. Sattlegger, and C. Steger, “Uninformed students: Student- teacher anomaly detection with discriminative latent embeddings,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

  18. [26]

    Fully convolutional cross-scale-flows for image-based defect detection,

    M. Rudolph, T. Wehrbein, B. Rosenhahn, and B. Wandt, “Fully convolutional cross-scale-flows for image-based defect detection,” in Winter Conference on Applications of Computer Vision (WACV), Jan. 2022. [Online]. Available: arxiv

  19. [27]

    Asymmetric student-teacher networks for industrial anomaly detection,

    ——, “Asymmetric student-teacher networks for industrial anomaly detection,” in Winter Conference on Applications of Computer Vision (WACV), Jan. 2023

  20. [28]

    Shape-guided dual-memory learning for 3d anomaly detection,

    Y .-M. Chu, C. Liu, T.-I. Hsieh, H.-T. Chen, and T.-L. Liu, “Shape-guided dual-memory learning for 3d anomaly detection,” in Proceedings of the 40th International Conference on Machine Learning, 2023, pp. 6185–6194

  21. [29]

    Looking 3d: Anomaly detection with 2d-3d alignment,

    A. K. Bhunia, C. Li, and H. Bilen, “Looking 3d: Anomaly detection with 2d-3d alignment,” CVPR, 2024

  22. [30]

    Fence theorem: Towards dual-objective semantic-structure isolation in preprocessing phase for 3d anomaly detection,

    H. Liang, J. Zhou, X. Chen, T. Dai, J. Wang, and C. Gao, “Fence theorem: Towards dual-objective semantic-structure isolation in preprocessing phase for 3d anomaly detection,”

  23. [31]

    Multimodal industrial anomaly detection via hybrid fusion,

    Y . Wang, J. Peng, J. Zhang, R. Yi, Y . Wang, and C. Wang, “Multimodal industrial anomaly detection via hybrid fusion,” 2023

  24. [32]

    Splatpose+: Real-time image-based pose-agnostic 3d anomaly detection,

    Y . Liu, Y . S. Hu, Y . Chen, and J. Zelek, “Splatpose+: Real-time image-based pose-agnostic 3d anomaly detection,” 2024. [Online]. Available: https://arxiv.org/abs/2410.12080

  25. [33]

    Splatpose & detect: Pose-agnostic 3d anomaly detection,

    M. Kruse, M. Rudolph, D. Woiwode, and B. Rosenhahn, “Splatpose & detect: Pose-agnostic 3d anomaly detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2024, pp. 3950–3960

  26. [34]

    Towards Generic Anomaly Detection and Understanding: Large-scale Visual-linguistic Model (GPT-4V) Takes the Lead,

    Y . Cao, X. Xu, C. Sun, X. Huang, and W. Shen, “Towards Generic Anomaly Detection and Understanding: Large-scale Visual-linguistic Model (GPT-4V) Takes the Lead,”arXiv e-prints, p. arXiv:2311.02782, Nov. 2023

  27. [35]

    Towards zero-shot 3d anomaly localization,

    Y . Wang, K.-C. Peng, and Y . Fu, “Towards zero-shot 3d anomaly localization,” inProceedings of the Winter Conference on Applications of Computer Vision (WACV) , February 2025, pp. 1447–1456

  28. [36]

    Pointad: Comprehending 3d anomalies from points and pixels for zero-shot 3d anomaly detection,

    Q. Zhou, J. Yan, S. He, W. Meng, and J. Chen, “Pointad: Comprehending 3d anomalies from points and pixels for zero-shot 3d anomaly detection,” in Advances in Neural Information Processing Systems , A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zha...

  29. [37]

    Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection,

    Y . Cao, J. Zhang, L. Frittoli, Y . Cheng, W. Shen, and G. Boracchi, “Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection,” inEuropean Conference on Computer Vision, 2024

  30. [38]

    Exploring the potential of encoder-free architectures in 3d lmms,

    Y . Tang, Z. Guo, Z. Wang, R. Zhang, Q. Chen, J. Liu, D. Qu, Z. Wang, D. Wang, X. Li, and B. Zhao, “Exploring the potential of encoder-free architectures in 3d lmms,” 2025. [Online]. Available: https://arxiv.org/abs/2502.09620

  31. [39]

    Towards zero-shot point cloud anomaly detection: A multi-view projection framework,

    Y . Cheng, Y . Cao, G. Xie, Z. Lu, and W. Shen, “Towards zero-shot point cloud anomaly detection: A multi-view projection framework,” 2024. [Online]. Available: https://arxiv.org/abs/2409.13162

  32. [40]

    3d-pnas: 3d industrial surface anomaly synthesis with perlin noise,

    Y . Cheng and J. Du, “3d-pnas: 3d industrial surface anomaly synthesis with perlin noise,” 2025. [Online]. Available: https://arxiv.org/abs/2504.12856

  33. [41]

    4d spatio-temporal convnets: Minkowski convolutional neural networks,

    C. Choy, J. Gwak, and S. Savarese, “4d spatio-temporal convnets: Minkowski convolutional neural networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 3075–3084

  34. [42]

    Surface defect detection methods for industrial products: A review,

    Y . Chen, Y . Ding, F. Zhao, E. Zhang, Z. Wu, and L. Shao, “Surface defect detection methods for industrial products: A review,” Applied Sciences, vol. 11, no. 16, 2021. [Online]. Available: https://www.mdpi.com/2076-3417/11/16/7657

  35. [43]

    The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization,

    P. Bergmann, X. Jin, D. Sattlegger, and C. Steger, “The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization,” in Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS ...

  36. [44]

    Complementary pseudo multimodal feature for point cloud anomaly detection,

    Y . Cao, X. Xu, and W. Shen, “Complementary pseudo multimodal feature for point cloud anomaly detection,” arXiv preprint arXiv:2303.13194, 2023

  37. [45]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    R. Q. Charles, H. Su, M. Kaichun, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 77–85

  38. [46]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” 2017. [Online]. Available: https://arxiv.org/abs/1706.02413

  39. [47]

    Point transformer,

    H. Zhao, L. Jiang, J. Jia, P. Torr, and V . Koltun, “Point transformer,” 2021. [Online]. Available: https://arxiv.org/abs/2012.09164

  40. [48]

    Dynamic graph cnn for learning on point clouds,

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,” ACM Transactions on Graphics (TOG), 2019

  41. [49]

    3d shapenets: A deep representation for volumetric shapes,

    Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao, “3d shapenets: A deep representation for volumetric shapes,” 2015. [Online]. Available: https://arxiv.org/abs/1406.5670

  42. [50]

    A scalable active framework for region annotation in 3d shape collections,

    L. Yi, V . G. Kim, D. Ceylan, I.-C. Shen, M. Yan, H. Su, C. Lu, Q. Huang, A. Sheffer, and L. Guibas, “A scalable active framework for region annotation in 3d shape collections,”SIG- GRAPH Asia, 2016

  43. [51]

    Sgdr: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” 2017. [Online]. Available: https://arxiv.org/abs/1608.03983

  44. [52]

    Boosting global-local feature matching via anomaly synthesis for multi-class point cloud anomaly detection,

    Y . Cheng, Y . Cao, D. Wang, W. Shen, and W. Li, “Boosting global-local feature matching via anomaly synthesis for multi-class point cloud anomaly detection,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 12 560–12 571, 2025

  45. [53]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling,

    X. Yu, L. Tang, Y . Rao, T. Huang, J. Zhou, and J. Lu, “Point-bert: Pre-training 3d point cloud transformers with masked point modeling,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022

  46. [54]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2015. [Online]. Available: https://arxiv.org/abs/1512.03385

  47. [55]

    A ransac-based approach to model fitting and its application to finding cylinders in range data,

    R. C. Bolles and M. A. Fischler, “A ransac-based approach to model fitting and its application to finding cylinders in range data,” in Proceedings of the 7th International Joint Conference on 26 Artificial Intelligence - Volume 2, ser. IJCAI’81. San Francisco, CA, USA: Morgan ...

  48. [56]

    Masked autoencoders for point cloud self-supervised learning,

    Y . Pang, W. Wang, F. E. Tay, W. Liu, Y . Tian, and L. Yuan, “Masked autoencoders for point cloud self-supervised learning,” in Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part II. Springer, 2022, pp. 604–621

  49. [57]

    Deep learning applied to computational mechanics: A comprehensive review, state of the art, and the classics,

    A. H. Loc Vu-Quoc, “Deep learning applied to computational mechanics: A comprehensive review, state of the art, and the classics,” Computer Modeling in Engineering & Sciences , vol. 137, no. 2, pp. 1069–1343, 2023. [Online]. Available: http://www.techscience.com/CMES/v137n2/53358

  50. [58]

    Multilayer feedforward networks are universal approximators,

    K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural Networks , vol. 2, no. 5, pp. 359–366, 1989. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0893608089900208

  51. [59]

    Reconstruct anything model: a lightweight foundation model for computational imaging,

    M. Terris, S. Hurault, M. Song, and J. Tachella, “Reconstruct anything model: a lightweight foundation model for computational imaging,” 2025. [Online]. Available: https://arxiv.org/abs/2503.08915

  52. [60]

    Dream to manipulate: Compositional world models empowering robot imitation learning with imagination,

    L. Barcellona, A. Zadaianchuk, D. Allegro, S. Papa, S. Ghidoni, and E. Gavves, “Dream to manipulate: Compositional world models empowering robot imitation learning with imagination,” in The Thirteenth International Conference on Learning Representations, 2025. [Online]. Availa...

  53. [2021]

    Available: https://arxiv.org/abs/2005.02357

    [Online]. Available: https://arxiv.org/abs/2005.02357

  54. [2025]

    Available: https://arxiv.org/abs/2503.01100

    [Online]. Available: https://arxiv.org/abs/2503.01100

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.