Pith. sign in

REVIEW 4 major objections 7 minor 44 references

C3D-AD: Toward Continual 3D Anomaly Detection via Kernel Attention with Learnable Advisor

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A point-cloud reconstruction model can serve many product classes and keep learning new ones without exemplar memory, by storing cross-task knowledge in a learnable attention advisor.

desk verdict First formulation of continual 3D anomaly detection with a plausible method and modest, consistent gains; the advisor-matrix memory mechanism is the weak point and needs a task-count stress test. read the letter →

arxiv 2508.01311 v1 pith:YSTD7KEL submitted 2025-08-02 cs.CV cs.LG

Haoquan Lu , Hanzhe Liang , Jie Zhang , Chenxi Hu , Jinbao Wang , Can Gao This is my paper
classification cs.CVcs.LG
keywords continuallearning3Danomalydetectionpointcloudsclass-incrementalkernelattentioncatastrophicforgettingreconstruction-basedindustrialinspection
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

This paper tries to establish that 3D anomaly detection can be made class-incremental: one model can keep detecting defects in previously seen product categories while learning to reconstruct new categories from normal samples alone, without storing past data. It proposes C3D-AD, which combines a random-feature kernel attention layer for extracting category-agnostic local features, a learnable attention advisor matrix inside the encoder and decoder that updates to absorb new-class information and drop redundant old information, and a perturbation-based reconstruction loss that keeps the model's outputs consistent across time. The paper reports average object-level AUROC of 66.4%, 83.1%, and 63.4% on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD, ahead of continual variants of previous 3D anomaly detection methods, and claims this is the first class-incremental 3D anomaly detection framework. A reader should care because industrial inspection systems meet new products continuously, and this design promises a single model that grows with the product line rather than being retrained from scratch.

What carries the argument

The load-bearing object is the learnable advisor matrix $S \in \mathbb{R}^{d \times m}$ in KAA, a continually updated attention cache that plays the role of the keys and values in linear kernel attention. Its update rule writes new class information into $S$ while subtracting a term said to remove redundant old information, and the attention output is $\phi(Q_t) S_t^\top$. KAL's positive random feature map $\phi$ creates the unified space in which all tasks' features are comparable, and RPP's worst-case perturbation loss constrains future outputs to stay close to current ones. The three modules together carry the continual learning claim: KAL prevents feature-space drift, KAA stores cross-task knowledge in one matrix, and RPP prevents reconstruction drift between tasks.

What would settle it

Train C3D-AD on Anomaly-ShapeNet split into ten sequential tasks instead of four, and measure object-level AUROC on the first task's classes after every new task; if that first-task score collapses toward 0.5 while later tasks stay strong, the advisor has not preserved the earliest classes.

Watch

Extended reading notes

Core claim

The central claim is that catastrophic forgetting in 3D anomaly detection can be avoided without rehearsal memory by unifying the feature space and compressing past knowledge into a small learned object. In C3D-AD, the KAL layer maps every task's point cloud groups through a fixed random-feature kernel, so different categories live in one normalized feature space. Inside the encoder–decoder, the KAA advisor, a matrix $S \in \mathbb{R}^{d \times m}$, is updated by $S_t = S_{t-1} - v^{\mathrm{old}}_t \phi_t(k)^\top + v^{\mathrm{new}}_t \phi_t(k)^\top$, which the paper reads as deleting redundant old associations while writing new ones; this same $S$ is the only stored record of previous tasks. The RPP loss then simulates the worst future parameter perturbation and asks the current model to reconstruct the same data identically after that perturbation, a constraint the paper formalizes with a generalization bound. Together these pieces are claimed to preserve old-category detection while learning new categories, and the reported AUROC averages support the claim on three datasets.

Load-bearing premise

The whole continual guarantee rests on one fixed-size advisor matrix having enough room to keep the detection-relevant information of every class it has seen, even though it never stores actual past point clouds.

Editorial extensions

If this is right

  • Deployed inspection systems could add new product categories by training only on that category's normal point clouds, with one model serving all categories seen so far.
  • No exemplar memory or raw data replay is needed, which helps when past production data cannot be stored or reused.
  • The linear $O(n)$ attention in KAL and KAA keeps runtime and memory growth predictable as point-cloud token counts rise.
  • The three public benchmarks with fixed task splits become a common evaluation ground for future class-incremental 3D anomaly detection methods.
  • The RPP worst-case perturbation constraint gives reconstruction-based continual learners a parameter-level regularization that does not depend on stored samples.

Reading between the lines

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

  • A fixed-size advisor implies a finite memory budget: the paper evaluates at most four tasks, so on longer category streams the matrix $S$ would likely need to grow, be split into per-class experts, or be refreshed; this scaling regime is left untested.
  • Because all three datasets use only a handful of tasks each, average AUROC may hide asymmetric forgetting; per-task retention curves on longer schedules would be the natural next experiment.
  • The same random-feature kernel attention design could transfer to continual point-cloud segmentation or object detection, where class-incremental drift is also a bottleneck.
  • The ablations suggest KAL and RPP contribute more than KAA, so an economical variant that relies on feature-space normalization plus perturbation rehearsal, without a separate advisor, is a plausible simplification worth testing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes C3D-AD, a framework for class-incremental 3D anomaly detection. It combines three components: a Kernel Attention with random feature Layer (KAL) that maps point-cloud features into a unified kernel space; a Kernel Attention with learnable Advisor (KAA) that maintains a fixed-size advisor matrix S inside the encoder-decoder and updates it online; and a Reconstruction with Parameter Perturbation (RPP) loss that regularizes the model against future output drift. The method is evaluated on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD under 3-4 task class-incremental protocols, reporting object-level mean AUROC and average ranking. The authors claim state-of-the-art results, including average AUROC of 66.4%, 83.1%, and 63.4% on the three datasets, and state that this is the first attempt at class-incremental 3D anomaly detection.

Significance. If the continual-learning claim is validated, the paper opens a useful direction: it addresses a realistic deployment scenario where new object classes appear over time, and it avoids storing exemplars. The ablations in Table 3 show that each proposed module contributes to the final result, and the complexity analysis in Figure 4 is a useful practical contribution. The strength of the paper is the clear formulation of the problem and the breadth of experiments across three datasets. However, the central memory mechanism is not convincingly validated: the evaluation lacks per-task forgetting metrics, longer task sequences, error bars or multiple seeds, and a direct capacity analysis of the advisor matrix S. These gaps are load-bearing because the continual-learning claim depends on S retaining old-class information.

major comments (4)
  1. [Kernel Attention with learnable Advisor (Eqs. 8-11)] The continual-learning claim rests on the advisor matrix S, but the update in Eq. (10) is an exponential-moving-average overwrite: for each current-task key, the old value v_old is replaced by v_new, and because tasks are disjoint, old-task keys are never revisited. With the default m=10, S spans a 10-dimensional subspace per layer, while Anomaly-ShapeNet has 40 classes, so the capacity of S to encode all past classes is asserted rather than demonstrated. Please provide per-task forgetting curves, a capacity sweep over m, and a longer task sequence (e.g., 10 tasks) to substantiate the memory claim.
  2. [Performance on Continual Anomaly Detection (Table 1)] The reported metric is a single mean AUROC evaluated on a cumulative test set at each step, not per-task AUROC or a standard continual-learning metric. Because the test set grows with each task, a decline such as Real3D-AD from 0.805 at task 1 to 0.664 at task 4 may reflect increased task difficulty rather than forgetting. Please report per-task AUROC after each training step, average accuracy, forgetting, and backward transfer, so that the continual-learning claim can be directly assessed.
  3. [Experiments (Tables 1-2)] All experimental results appear to be from a single run, with no error bars or multiple seeds. The headline improvements over Continual-MC3D-AD are only 3.9, 3.4, and 3.0 percentage points in Table 1, and the advantage over MC3D-AD in Table 2 is 0.4 points. Given the small margins and the high variance typical of anomaly-detection benchmarks, please report means and standard deviations over at least three seeds, and state the number of runs in the implementation details.
  4. [Generalization Error Bound of RPP Loss (Theorem 1, Eq. 14)] Theorem 1 is stated without a proof and without definitions of the quantities R(psi) and L_theta; the text refers to supplemental materials that are not included in the submission. Since the theorem is used to justify the choice of epsilon in Eq. (13), please either provide a complete proof with all regularity conditions in the appendix or remove the theoretical claim and present the epsilon sensitivity as an empirical observation.
minor comments (7)
  1. [Kernel Attention with learnable Advisor (Eq. 8)] There is a typo in the gradient expression: the second term should be (1+alpha) v phi(k)^T, not (1+alpha) phi(k) v^T; please also state explicitly that phi(x) is a column vector in R^m and S is in R^{d x m}.
  2. [Throughout] The name is inconsistent: the abstract and method sections use "C3D-AD" while the framework overview uses "C3DAD"; please standardize.
  3. [Figure 2] The caption contains a garbled line "Tasks AUROC 0 4321" that should be removed or replaced with a clear description of the example task schedule.
  4. [Table 2] The column headers for Anomaly-ShapeNet categories are truncated and hard to read; consider rotating the table, using full class names, or grouping the classes by task.
  5. [Implementation Details] The sentence "The sample scaling factor eta is 10 to cover the whole point cloud" is vague; please report the sensitivity of eta or justify the chosen value in relation to Eq. (2).
  6. [Table 5] The m-sweep reports only AUROC and efficiency; since m is the width of the advisor memory, please also report per-task retention or forgetting to show whether larger m helps memory.
  7. [Preliminary notation (Eqs. 1-2)] In Eq. (2), the index i in \bar{p}_i is used both for the center point whose neighborhood is considered and inside the average over all centers; please clarify the notation, e.g., by writing \bar{p}_c for the anchor center.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: KAA is an explicit gradient-based update for a per-task objective, RPP is an independent regularizer, and all reported AUROC values are external benchmark evaluations.

full rationale

The paper's central continual-learning mechanism is KAA, whose advisor S is updated by Eq. (9)-(10) as the gradient step of the explicit objective in Eq. (8), min_S ||Sφ(k)-v||^2 - αTr(v^T Sφ(k)). This is a concrete per-task optimization, not a quantity fitted to the target AUROC numbers. The RPP loss in Eq. (13) is a perturbation-based consistency regularizer evaluated on current data, and Theorem 1 gives a generalization bound rather than encoding the benchmark result. The reported performances are measured on held-out anomalous and normal samples from public datasets under a defined task-incremental protocol, so the results are external evidence rather than consequences of the definitions. The paper does cite prior work by some of the same authors (Liang et al. 2025a,b; Cheng et al./MC3D-AD), but these citations are contextual background for reconstruction-based and multiclass AD, and the continual-learning claim does not rest on them. The concern that the fixed-size advisor may not retain many old classes is an assumption about model capacity and update effectiveness, not a circular step; it would be a correctness or robustness issue, not a derivation that reduces to its own inputs. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. Therefore the derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the listed hand-tuned hyperparameters, unproved background results, and design assumptions. The paper introduces no new physical entities; the advisor S is a trainable matrix, not an invented entity in the sense of a new particle or force.

free parameters (5)
  • alpha (direction alignment) = 0.7
    Hyperparameter for the KAA objective in Eq. (8), chosen via sensitivity analysis in Figure 3.
  • beta (advisor learning rate) = 0.7
    Hyperparameter for the KAA update in Eq. (9), chosen via sensitivity analysis in Figure 3.
  • epsilon (perturbation bound) = scanned in [0.01, 10]
    Bound on the RPP perturbation delta in Eq. (13), scanned in Table 4 and linked to the generalization bound in Theorem 1.
  • eta (radius scaling factor) = 10
    Scaling factor for the adaptive neighborhood radius in Eq. (2), set to cover the whole point cloud.
  • m (number of random features) = 10 (default), up to 1000 in analysis
    Dimension of the random feature map phi in Eq. (7), affecting accuracy and efficiency as shown in Table 5.
assumptions (4)
  • standard math Positive Random Features approximate the softmax kernel with finite-dimensional feature maps
    KAL relies on Eq. (7) from Choromanski et al. 2021 to map tokens into a unified kernel space with linear complexity.
  • domain assumption Reconstruction error between input and reconstructed point cloud tokens is a valid anomaly score
    The whole pipeline detects anomalies by comparing feature tokens and reconstruction tokens, an assumption common to reconstruction-based anomaly detection methods.
  • ad hoc to paper The KAA advisor objective in Eq. (8) yields an update rule that preserves old knowledge while learning new knowledge
    Eq. (10) is a Hebbian-style update, but the paper provides no proof that the fixed-size S actually retains discriminative information for all past tasks.
  • ad hoc to paper The worst-case parameter perturbation in Eq. (13) approximates the model's future output drift
    RPP assumes that optimizing against a bounded random perturbation simulates future task updates, which is an unverified modeling choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of C3D-AD: Toward Continual 3D Anomaly Detection via Kernel Attention with Learnable Advisor." pith.science (2026). https://pith.science/paper/YSTD7KEL

@misc{pith2026250801311,
  author       = {Pith},
  title        = {Pith review of: C3D-AD: Toward Continual 3D Anomaly Detection via Kernel Attention with Learnable Advisor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YSTD7KEL}},
  note         = {Machine review of arXiv:2508.01311}
}
read the original abstract

3D Anomaly Detection (AD) has shown great potential in detecting anomalies or defects of high-precision industrial products. However, existing methods are typically trained in a class-specific manner and also lack the capability of learning from emerging classes. In this study, we proposed a continual learning framework named Continual 3D Anomaly Detection (C3D-AD), which can not only learn generalized representations for multi-class point clouds but also handle new classes emerging over time.Specifically, in the feature extraction module, to extract generalized local features from diverse product types of different tasks efficiently, Kernel Attention with random feature Layer (KAL) is introduced, which normalizes the feature space. Then, to reconstruct data correctly and continually, an efficient Kernel Attention with learnable Advisor (KAA) mechanism is proposed, which learns the information from new categories while discarding redundant old information within both the encoder and decoder. Finally, to keep the representation consistency over tasks, a Reconstruction with Parameter Perturbation (RPP) module is proposed by designing a representation rehearsal loss function, which ensures that the model remembers previous category information and returns category-adaptive representation.Extensive experiments on three public datasets demonstrate the effectiveness of the proposed method, achieving an average performance of 66.4%, 83.1%, and 63.4% AUROC on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD, respectively.

Figures

Figures reproduced from arXiv: 2508.01311 by the authors.

Figure 1
Figure 1. Difference between class-specific model and C3D [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of C3D-AD. The training point cloud data is aggregated into groups according to the center points. Feature tokens are generated by extracting features from both group centers and point groups, utilizing the Kernel Attention with random feature Layer (KAL) module in linear complexity. Then, feature tokens are input into the Encoder-Decoder, employing Kernel Attention with learnable Advisor (KAA) mechanis… view at source ↗
Figure 3
Figure 3. Sensitivity of α and β on Anomaly-ShapeNet. in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Inference time (s) and memory usage (GB) vs. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 34 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Memory Efficient Continual Learning for Edge-Based Visual Anomaly Detection

    Barusco, M.; D'Antoni, L.; Pezze, D. D.; Borsatti, F.; and Susto, G. A. 2025. Memory Efficient Continual Learning for Edge-Based Visual Anomaly Detection. arXiv:2503.02691

    work page Pith review arXiv 2025

  4. [4]

    Batra, H.; and Clark, R. 2024. EVCL: Elastic Variational Continual Learning with Weight Consolidation. arXiv:2406.15972

    arXiv 2024

  5. [5]

    Bergmann, P.; and Sattlegger, D. 2023. Anomaly Detection in 3D Point Clouds Using Deep Geometric Descriptors. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2613--2623

    work page 2023

  6. [6]

    Bonicelli, L.; Boschini, M.; Porrello, A.; Spampinato, C.; and Calderara, S. 2022. On the Effectiveness of Lipschitz-Driven Rehearsal in Continual Learning. In Oh, A. H.; Agarwal, A.; Belgrave, D.; and Cho, K., eds., Advances in Neural Information Processing Systems

    work page 2022

  7. [7]

    Cao, Y.; Xu, X.; and Shen, W. 2024. Complementary pseudo multimodal feature for point cloud anomaly detection. Pattern Recognition(PR), 156: 110761

    work page 2024

  8. [8]

    Q.; Su, H.; Kaichun, M.; and Guibas, L

    Charles, R. Q.; Su, H.; Kaichun, M.; and Guibas, L. J. 2017. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 77--85

    work page 2017

Show all 44 references
  1. [9]

    Chen, R.; Xie, G.; Liu, J.; Wang, J.; Luo, Z.; Wang, J.; and Zheng, F. 2023. EasyNet: An Easy Network for 3D Industrial Anomaly Detection. arXiv:2307.13925

    2023 arXiv

  2. [10]

    Cheng, J.; Gao, C.; Zhou, J.; Wen, J.; Dai, T.; and Wang, J. 2025. MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection. arXiv:2505.01969

    2025

  3. [11]

    M.; Likhosherstov, V.; Dohan, D.; Song, X.; Gane, A.; Sarlos, T.; Hawkins, P.; Davis, J

    Choromanski, K. M.; Likhosherstov, V.; Dohan, D.; Song, X.; Gane, A.; Sarlos, T.; Hawkins, P.; Davis, J. Q.; Mohiuddin, A.; Kaiser, L.; Belanger, D. B.; Colwell, L. J.; and Weller, A. 2021. Rethinking Attention with Performers. In International Conference on Learning Representations

    2021

  4. [12]

    Eskandar, M.; Imtiaz, T.; Hill, D.; Wang, Z.; and Dy, J. 2025. STAR : Stability-Inducing Weight Perturbation for Continual Learning. In The Thirteenth International Conference on Learning Representations

    2025

  5. [13]

    Gu, Z.; Zhang, J.; Liu, L.; Chen, X.; Peng, J.; Gan, Z.; Jiang, G.; Shu, A.; Wang, Y.; and Ma, L. 2024. Rethinking Reverse Distillation for Multi-Modal Anomaly Detection. Proceedings of the AAAI Conference on Artificial Intelligence, 38(8): 8445--8453

    2024

  6. [14]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising Diffusion Probabilistic Models. In Larochelle, H.; Ranzato, M.; Hadsell, R.; Balcan, M.; and Lin, H., eds., Advances in Neural Information Processing Systems, volume 33, 6840--6851. Curran Associates, Inc

    2020

  7. [15]

    Horwitz, E.; and Hoshen, Y. 2023. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2968--2977

    2023

  8. [16]

    Hu, L.; Gan, Z.; Deng, L.; Liang, J.; Liang, L.; Huang, S.; and Chen, T. 2025. ReplayCAD: Generative Diffusion Replay for Continual Anomaly Detection. arXiv preprint arXiv:2505.06603

    2025 arXiv

  9. [17]

    P.; and Welling, M

    Kingma, D. P.; and Welling, M. 2022. Auto-Encoding Variational Bayes. arXiv:1312.6114

    2022 arXiv

  10. [18]

    Kruse, M.; Rudolph, M.; Woiwode, D.; and Rosenhahn, B. 2024. SplatPose & Detect: Pose-Agnostic 3D Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 3950--3960

    2024

  11. [19]

    Li, W.; Xu, X.; Gu, Y.; Zheng, B.; Gao, S.; and Wu, Y. 2024. 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 Recognitio...

    2024

  12. [20]

    Li, W.; Zhan, J.; Wang, J.; Xia, B.; Gao, B.-B.; Liu, J.; Wang, C.; and Zheng, F. 2022. Towards Continual Adaptation in Industrial Anomaly Detection. In Proceedings of the 30th ACM International Conference on Multimedia, 2871–2880. Association for Computing Machinery

    2022

  13. [21]

    Li, W.; Zheng, B.; Xu, X.; Gan, J.; Lu, F.; Li, X.; Ni, N.; Tian, Z.; Huang, X.; Gao, S.; and Wu, Y. 2025 a . Multi-Sensor Object Anomaly Detection: Unifying Appearance, Geometry, and Internal Properties. In Proceedings of the Computer Vision and Pattern Recognition Conference...

    2025

  14. [22]

    Li, X.; Tan, X.; Chen, Z.; Zhang, Z.; Zhang, R.; Guo, R.; Jiang, G.; Chen, Y.; Qu, Y.; Ma, L.; and Xie, Y. 2025 b . One-for-More: Continual Diffusion Model for Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 4766--4775

    2025

  15. [23]

    Liang, H.; Xie, G.; Hou, C.; Wang, B.; Gao, C.; and Wang, J. 2025 a . Look Inside for More: Internal Spatial Modality Perception for 3D Anomaly Detection. Proceedings of the AAAI Conference on Artificial Intelligence, 39(5): 5146--5154

    2025

  16. [24]

    Liang, H.; Zhang, J.; Dai, T.; Shen, L.; Wang, J.; and Gao, C. 2025 b . Taming Anomalies with Down-Up Sampling Networks: Group Center Preserving Reconstruction for 3D Anomaly Detection. arXiv:2507.03903

    2025

  17. [25]

    Liu, J.; Wu, K.; Nie, Q.; Chen, Y.; Gao, B.-B.; Liu, Y.; Wang, J.; Wang, C.; and Zheng, F. 2024 a . Unsupervised Continual Anomaly Detection with Contrastively-Learned Prompt. Proceedings of the AAAI Conference on Artificial Intelligence, 38(4): 3639--3647

    2024

  18. [26]

    Liu, J.; Xie, G.; Chen, R.; Li, X.; Wang, J.; Liu, Y.; Wang, C.; and Zheng, F. 2023. Real3D-AD: A Dataset of Point Cloud Anomaly Detection. In Oh, A.; Naumann, T.; Globerson, A.; Saenko, K.; Hardt, M.; and Levine, S., eds., Advances in Neural Information Processing Systems, vo...

    2023

  19. [27]

    Liu, J.; Xie, G.; Wang, J.; Li, S.; Wang, C.; Zheng, F.; and Jin, Y. 2024 b . Deep industrial image anomaly detection: A survey. Machine Intelligence Research, 21(1): 104--135

    2024

  20. [28]

    Mallya, A.; Davis, D.; and Lazebnik, S. 2018. Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights. In Proceedings of the European Conference on Computer Vision (ECCV)

    2018

  21. [29]

    Mallya, A.; and Lazebnik, S. 2018. PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning. arXiv:1711.05769

    2018 arXiv

  22. [30]

    Masuda, M.; Hachiuma, R.; Fujii, R.; Saito, H.; and Sekikawa, Y. 2021. Toward Unsupervised 3d Point Cloud Anomaly Detection Using Variational Autoencoder. In 2021 IEEE International Conference on Image Processing (ICIP), 3118--3122

    2021

  23. [31]

    E.; Liu, W.; Tian, Y.; and Yuan, L

    Pang, Y.; Wang, W.; Tay, F. E.; Liu, W.; Tian, Y.; and Yuan, L. 2022. 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, 604--621. Springer

    2022

  24. [32]

    Qin, J.; Gu, C.; Yu, J.; and Zhang, C. 2023. Teacher–student network for 3D point cloud anomaly detection with few normal samples. Expert Systems with Applications, 228: 120371

    2023

  25. [33]

    Rebuffi, S.-A.; Kolesnikov, A.; Sperl, G.; and Lampert, C. H. 2017. iCaRL: Incremental Classifier and Representation Learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2017

  26. [34]

    Riemer, M.; Klinger, T.; Bouneffouf, D.; and Franceschini, M. 2019. Scalable recollections for continual lifelong learning. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Confere...

    2019

  27. [35]

    Roth, K.; Pemula, L.; Zepeda, J.; Schölkopf, B.; Brox, T.; and Gehler, P. 2022. Towards Total Recall in Industrial Anomaly Detection. arXiv:2106.08265

    2022 arXiv

  28. [36]

    Rudolph, M.; Wehrbein, T.; Rosenhahn, B.; and Wandt, B. 2023. Asymmetric Student-Teacher Networks for Industrial Anomaly Detection. In Winter Conference on Applications of Computer Vision (WACV)

    2023

  29. [37]

    W.; Ge, M.; Chen, Y.-C.; and Tsung, F

    Tang, J.; Lu, H.; Xu, X.; Wu, R.; Hu, S.; Zhang, T.; Cheng, T. W.; Ge, M.; Chen, Y.-C.; and Tsung, F. 2025. An Incremental Unified Framework for Small Defect Inspection. In Computer Vision -- ECCV 2024, 307--324

    2025

  30. [38]

    Tu, Y.; Zhang, B.; Liu, L.; Li, Y.; Zhang, J.; Wang, Y.; Wang, C.; and Zhao, C. 2025. Self-supervised Feature Adaptation for 3D Industrial Anomaly Detection. In Computer Vision -- ECCV 2024, 75--91

    2025

  31. [39]

    Wang, J.; Xu, H.; Chen, X.; Xu, H.; Huang, Y.; Ding, X.; and Tu, X. 2025. Exploiting Point-Language Models with Dual-Prompts for 3D Anomaly Detection. arXiv:2502.11307

    2025 arXiv

  32. [40]

    Wang, Y.; Peng, J.; Zhang, J.; Yi, R.; Wang, Y.; and Wang, C. 2023. Multimodal Industrial Anomaly Detection via Hybrid Fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8032--8041

    2023

  33. [41]

    Wang, Z.; Jian, T.; Chowdhury, K.; Wang, Y.; Dy, J.; and Ioannidis, S. 2020. Learn-Prune-Share for Lifelong Learning. In 2020 IEEE International Conference on Data Mining (ICDM), 641--650

    2020

  34. [42]

    Wu, Z.; Song, S.; Khosla, A.; Yu, F.; Zhang, L.; Tang, X.; and Xiao, J. 2015. 3D ShapeNets: A Deep Representation for Volumetric Shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2015

  35. [43]

    Ye, J.; Zhao, W.; Yang, X.; Cheng, G.; and Huang, K. 2024. PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection. arXiv:2412.12617

    2024 arXiv

  36. [44]

    Zhou, Z.; Wang, L.; Fang, N.; Wang, Z.; Qiu, L.; and Zhang, S. 2025. R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection. In Computer Vision -- ECCV 2024, 91--107

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.