Pith. sign in

REVIEW 3 major objections 5 minor 34 references

IAENet: An Importance-Aware Ensemble Model for 3D Point Cloud-Based Anomaly Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read IAENet claims to improve 3D point-cloud anomaly localization by adaptively reweighting a 2D pretrained expert and a 3D expert, reaching AUPRO 0.944.

desk verdict A neat fusion idea, backed by solid ablations, but the missing point-to-grid mapping makes the headline localization numbers unverifiable as reported. read the letter →

arxiv 2508.20492 v1 pith:GH627I5G submitted 2025-08-28 cs.CV

Xuanming Cao , Chengyu Tao , Yifeng Cheng , Juan Du This is my paper
classification cs.CV
keywords 3Dpointcloudanomalydetectionensemblefusionimportance-awarememorybankMVTec3D-ADsurfaceinspectiondepthprojection
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

The paper proposes IAENet, an unsupervised ensemble for 3D point-cloud surface anomaly detection. Its central claim is that the lack of powerful pretrained 3D backbones can be sidestepped by pairing a frozen 2D foundation model, fed with rendered depth images, with a dedicated 3D geometric expert, and then learning how much to trust each expert at every spatial location rather than combining scores by a fixed rule like max or addition. The proposed Importance-Aware Fusion module reweights the two anomaly score maps through a selector network before a nonlinear predictor produces the final per-point score. On MVTec 3D-AD the paper reports a new best point-level localization result, AUPRO 0.944 versus 0.929 for prior best methods, with notably lower anomaly scores on normal regions and therefore a lower false-positive rate, which matters in industrial inspection.

What carries the argument

The central object is the Importance-Aware Fusion (IAF) module, made of a selector network and a predictor network. The selector takes the two expert score maps concatenated pixel-wise and outputs per-location importance weights; the predictor multiplies each map by its weight and applies a nonlinear mapping to produce the final fused score map. The module is trained on synthetic Cut-Paste anomalies with a predictor cross-entropy loss and a selector loss that combines a margin-based performance gate (which closes once fusion beats the better single expert) with an entropy regularizer that drives exploration from equal weights to specialization. Around this sits a dual memory bank built by sh

What would settle it

Re-implement the per-point evaluation by mapping each of the M points to its projected pixel in the depth image, assigning the fused score of that pixel with an explicit rule for out-of-image points and for pixels shared by many points, then recompute mean AUPRO on MVTec 3D-AD. If alternative tie-handling rules change AUPRO by more than a small margin, or produce visible misalignment, the reported gain over prior methods would not be a stable property of the method.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that decision-level fusion of 2D and 3D anomaly experts can be made reliable by learning per-location importance weights instead of assuming equal contributions. The 2D expert (pretrained Wide-ResNet-50-2 on a single frontal normal-based depth rendering, plus FPFH descriptors) tends to catch fine surface details but marks many normal points; the 3D expert (PointNet with SDF reconstruction) captures global geometry but can miss subtle defects. IAENet trains a selector network to assign importance scores to the two score maps, uses a margin-based selector loss with an entropy regularizer to force exploration of weight combinations and then stop when fusion be

Load-bearing premise

The localization result assumes that the fused h x w anomaly map on the rendered depth grid can be mapped back to the original point cloud without losing or misaligning points, but the paper does not specify this per-point mapping.

Editorial extensions

If this is right

  • If the reported numbers hold, learned importance-aware fusion outperforms fixed rules: on MVTec 3D-AD it beats max, addition, and OCSVM fusion on all three metrics by clear margins.
  • The 2D expert needs only a single frontal projection to be competitive, so the method avoids tuning multi-view projection counts and angles used by prior approaches, simplifying deployment.
  • Because the selector can down-weight a failing expert, the fused object-level score stays high even when one expert collapses, for example Cable gland with 2D expert 0.969, 3D expert 0.710, and fused 0.954.
  • The score compression on normal points should translate to fewer false alarms at operating points with strict false-positive budgets, which is the regime the AUPRO curves at small integration limits are meant to capture.

Reading between the lines

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

  • The IAF design is agnostic to what generates the score maps, so the same learned fusion could combine other pairs of detectors, such as RGB and point cloud or flow and memory-bank models; swapping in different experts on MVTec 3D-AD would be a direct test of whether the fusion mechanism, not the specific experts, drives the gain.
  • Since training uses synthetic Cut-Paste anomalies only, the paper leaves open how the learned importance weights behave on real defect types whose geometry differs from pasted patches; a per-anomaly-type breakdown of AUPRO would show whether the low false-positive rate holds for every defect class.
  • The paper's reported point-level metrics are computed on the original point cloud, while fusion happens on an h x w depth-image grid; spelling out the per-point mapping, and handling of points sharing a pixel or falling outside the rendered view, is a needed reproducibility check before the 0.944 AUPRO claim can be taken at face value.
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

3 major / 5 minor

Summary. The paper proposes IAENet, an ensemble framework for 3D point-cloud anomaly detection that combines a 3D expert (PointNet–SDF with dual memory-bank retrieval, following Shape-Guided) and a 2D expert (frozen ImageNet-pretrained ResNet on a single rendered depth map, with FPFH features) through a learned Importance-Aware Fusion (IAF) module. The IAF contains a selector network that reweights the two expert score maps per spatial location and a predictor network that nonlinearly maps the weighted maps to a final anomaly map. The fusion is trained on synthetically generated Cut-Paste anomalies using a margin-based selector loss with an entropy regularizer and a separate predictor loss. Experiments on MVTec 3D-AD report new state-of-the-art point-level localization results (P-AUROC 0.982, AUPRO@30% 0.944) and competitive object-level detection (O-AUROC 0.944), together with ablations of the fusion components and comparisons to max, addition, and OCSVM fusion.

Significance. If the results hold, the paper makes a useful practical contribution: it shows that a well-chosen 2D pretrained backbone can be adapted to 3D anomaly detection via a single depth projection, and that a learned, importance-aware fusion rule can outperform simple max/addition/OCSVM fusion. The proposed margin-gated selector loss is a novel mechanism for suppressing the influence of a poorly performing expert. The paper provides per-category results on a standard benchmark, AUPRO curves at multiple integration limits, and an ablation of the fusion components. However, the central localization claim is currently not verifiable because the paper never specifies how the grid-based fused anomaly map is converted to per-point scores, or how points outside the single frontal view are handled. Without this mapping, the reported AUPRO and false-positive-rate improvements cannot be reliably compared with prior methods that score the full point cloud.

major comments (3)
  1. [§3.3.1, §4.1.2, §4.1.3] The fusion module operates on h×w grids X1, X2 and outputs a grid A, but the point-level evaluation metrics P-AUROC and AUPRO are defined on the original M-point cloud. The paper never defines how per-point expert scores are rasterized into the depth-image grid, nor how the fused grid A is unprojected to per-point scores. It also does not state what happens to points that are not visible in the single frontal view (Section 4.1.2: 'we only project a frontal view') or that are occluded. If these points receive a default score of 0 or are excluded from evaluation, the reported AUPRO of 0.944 and the 'markedly lower false positive rate' claim are not comparable to baselines that score the full point cloud. The synthetic labels Y_i used in Eq. (18) must also be aligned to the grid; this alignment is not specified. Please provide the exact mapping in both directions, including handling of occl
  2. [§4.1.2] The 2D expert uses a single frontal depth projection, so it has no information about backside or occluded surface regions. Since the fusion is performed on the grid, it is unclear how the IAF module obtains a fused score for points that do not project into this view. If those points are assigned zero score by default, the point-level localization scores are inflated because many normal points would trivially produce zero false positives. The authors should report the fraction of points covered by the frontal view per category and describe how uncovered points are scored during both training and inference. This is essential for interpreting the comparison with CPMF and Shape-Guided, which score the full cloud.
  3. [§3.3.2, Eqs. (11)–(13), (18)–(19)] The loss is written as cross-entropy with a sum over i=1..c and c=1. For binary anomaly labels (y∈{0,1}), standard binary cross-entropy is -[y log p + (1-y) log(1-p)]. The formulation as written includes only the positive-class term, which would not penalize high anomaly scores on normal pixels and contradicts the paper's stated goal of suppressing normal-region scores. Please clarify the notation (e.g., sum over both classes) or present the actual binary cross-entropy used. This affects both the selector loss L_s and the predictor loss L_p.
minor comments (5)
  1. [Table 5] The component names in the ablation table appear swapped: the row labeled 'Predictor Network S_theta' and the row labeled 'Selector Network f_phi_i' contradict the definitions in Section 3.3.1 and Figure 4, where S_theta is the selector and f_phi is the predictor. Please relabel the rows/columns to match the notation used in the text.
  2. [Abstract and Table 4] There are typos: 'We introduce an novel' in the abstract, and 'AURPO' in Table 4 and Figure 7 should be 'AUPRO'.
  3. [§3.3.2, Eq. (16)–(18)] The sign of the entropy term is inconsistent across equations: Eq. (16) uses H = -S log S, but the expanded form in Eq. (18) is written with an outer minus sign and S log S. Please align the notation so the gradient direction is unambiguous.
  4. [§4.1.2] The statement that margin m=0.1 and λ=1 'may require fine-tuning for some categories' raises the question of whether per-category tuning was performed. If so, please report which categories were tuned and the values used; if not, clarify that a single global setting was used.
  5. [References [29]] The Cut-Paste synthetic data generation is attributed to the authors' own G2SF-MIAD paper. Please clarify the relationship and whether the synthetic data generation code/configuration will be released to facilitate reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: IAENet's fusion is trained on synthetic labels and evaluated on held-out real anomalies; self-citations are not load-bearing.

full rationale

The paper's central derivation is not circular. The IAF module is trained on a synthetic dataset D' = {(P'_i, Y_i)} generated by Cut-Paste, with per-point binary labels, and is evaluated on the real MVTec 3D-AD test set; the reported O-AUROC/P-AUROC/AUPRO numbers are empirical comparisons against external baselines, not consequences of the training objective. The margin loss (Eqs. 7-18) is a training signal: b is the minimum of the two expert cross-entropy losses computed on the fixed training maps, and the selector/predictor must actually generalize to unseen test point clouds to improve the metrics. Nothing in the definition forces the test-time AUPRO to be high. The one notable self-citation is [29] (G2SF-MIAD) for the Cut-Paste synthesis, but the paper itself describes the paste procedure (mask creation, cutting, pasting, random transformations), and the synthetic target is an independent training signal rather than an already-known prediction. The grid-to-point mapping between Sec. 3.3 (X1,X2 in R^{h x w}, A in R^{h x w}) and the point-level metrics of Sec. 4.1.3 (A in R^M) is never specified, which is a serious reproducibility/validity concern, but it is not circular: it does not make a claimed prediction equivalent to an input. The method is self-contained against external benchmarks, and the small self-citations in the intro/survey are not load-bearing. No circular step can be quoted or exhibited.

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

The central claim rests on reused experts and a learned fusion. The main contributions are the IAF network and its losses, which introduce hand-chosen hyperparameters (margin, lambda, synthetic data size) and domain assumptions about transfer from synthetic anomalies and about the validity of 2D features on depth maps.

free parameters (5)
  • margin m = 0.1
    Gates the selector loss; paper states it may require fine-tuning for some categories.
  • loss balance lambda = 1.0
    Weights the selector loss in the total loss; chosen by hand.
  • synthetic samples per class = 800
    Number of Cut-Paste samples generated for IAF training; no sweep reported.
  • nearest neighbor counts k1, k2 = not stated
    Retrieval neighborhood sizes for 3D and 2D memory banks; not specified in the paper.
  • learning rate = 0.01
    AdamW with cosine schedule; standard but manually set.
assumptions (4)
  • domain assumption PointNet-SDF pretrained on normal scans yields per-point anomaly scores via |s|.
    Adopted from Shape-Guided [11]; the entire 3D expert is reused without modification.
  • domain assumption ImageNet-pretrained ResNet features on a single textured depth map are informative for 3D anomaly detection.
    The 2D expert relies on this transfer; no analysis of projection angle or depth-texture choice is provided.
  • domain assumption Cut-Paste synthetic anomalies are a valid training proxy for real anomalies.
    IAF is trained exclusively on synthetic samples; generalization to real defects is assumed and only tested on MVTec 3D-AD.
  • domain assumption Shape-guided memory bank retention preserves the most discriminative normal features.
    Inherited from [11]; the paper adopts it without independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IAENet: An Importance-Aware Ensemble Model for 3D Point Cloud-Based Anomaly Detection." pith.science (2026). https://pith.science/paper/GH627I5G

@misc{pith2026250820492,
  author       = {Pith},
  title        = {Pith review of: IAENet: An Importance-Aware Ensemble Model for 3D Point Cloud-Based Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GH627I5G}},
  note         = {Machine review of arXiv:2508.20492}
}
read the original abstract

Surface anomaly detection is pivotal for ensuring product quality in industrial manufacturing. While 2D image-based methods have achieved remarkable success, 3D point cloud-based detection remains underexplored despite its richer geometric cues. We argue that the key bottleneck is the absence of powerful pretrained foundation backbones in 3D comparable to those in 2D. To bridge this gap, we propose Importance-Aware Ensemble Network (IAENet), an ensemble framework that synergizes 2D pretrained expert with 3D expert models. However, naively fusing predictions from disparate sources is non-trivial: existing strategies can be affected by a poorly performing modality and thus degrade overall accuracy. To address this challenge, We introduce an novel Importance-Aware Fusion (IAF) module that dynamically assesses the contribution of each source and reweights their anomaly scores. Furthermore, we devise critical loss functions that explicitly guide the optimization of IAF, enabling it to combine the collective knowledge of the source experts but also preserve their unique strengths, thereby enhancing the overall performance of anomaly detection. Extensive experiments on MVTec 3D-AD demonstrate that our IAENet achieves a new state-of-the-art with a markedly lower false positive rate, underscoring its practical value for industrial deployment.

Figures

Figures reproduced from arXiv: 2508.20492 by the authors.

Figure 1
Figure 1. The illustration of the anomaly score maps of different expert models on two representative objects from MVTec 3D-AD [5]. Despite receiving the same point cloud input, the 2D and 3D experts exhibit complementary strengths. When an anomaly is detected by one expert but not by another, our IAENet integrates both insights to accurately identify all anomalies. It also effectively suppresses the anomaly scores of normal … view at source ↗
Figure 2
Figure 2. The framework of our IAENet. During training, normal samples first pass through an anomaly synthesis module that generates pseudo anomalies (disabled at inference). The resulting anomalous point clouds are then processed by two source experts, yielding anomaly score maps 𝐗𝟏 and 𝐗𝟐 , respectively. Finally, the Importance-Aware Fusion (IAF) module adaptively reweights these scores and outputs the refined anomaly map 𝐀… view at source ↗
Figure 3
Figure 3. Illustration of the construction of shape-guided dual memory bank. textures, offering richer information than simple depth pro￾jection methods. This textured depth map is then processed by the pretrained ResNet to extract global semantic features. Finally, the global representation is concatenated with per￾point FPFH descriptors to yield the composite feature 𝐟2 ∈ ℝ𝑑2 that encodes both global context and local geome… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The architecture of proposed selector network and predictor network [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of anomaly score maps on the MVTec 3D-AD dataset. This is the visualization of the final output of the anomaly score map for each methodology, where blue represents values close to 0 and red represents larger anomaly score values [PITH_FULL_IMAG…
Figure 7
Figure 7. Figure 7: Anomaly localization performance (AURPO) of our method with comparison methods at different integration limits. this section, we will quantify and analyze this advantage by examining the results of AUPRO at different integration limits and the distribution of anomaly s…
Figure 6
Figure 6. Figure 6: Object-level (O-AUROC) and point-level (AUPRO) anomaly detection results on the MVTec 3D-AD dataset by the 2D expert, 3D expert, and our IAENet. accurately locates subtle anomalies with high confidence but also accurately discerns complex normal surfaces. 4.3. Analysis…
Figure 8
Figure 8. Figure 8: Point-level anomaly score distributions across three categories. The x-axis represents anomaly scores, and the y￾axis represents probability density. It can be observed that our approach effectively suppresses the anomaly scores in normal regions compared to CPMF. inte…
Figure 9
Figure 9. Figure 9: Qualitative comparison of different fusion strategies in representative categories, including Cable gland, Cookie, Peach and Foam. The left side shows the input point clouds, the fusion inputs and ground truth, and the right side shows the outputs of different fusion s…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 29 canonical work pages

  1. [1]

    Ammar,M.B.,Mendoza,A.,Belkhir,N.,Manzanera,A.,Franchi,G.,

    work page

  2. [2]

    Pni: industrial anomaly detection using position and neighborhood information, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Bae, J., Lee, J.H., Kim, S., 2023. Pni: industrial anomaly detection using position and neighborhood information, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6373– 6383

    work page 2023

  3. [3]

    Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C., 2019. Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9592–9600

    work page 2019

  4. [4]

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C., 2020. Unin- formedstudents:Student-teacheranomalydetectionwithdiscrimina- tive latent embeddings, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 4183–4192

    work page 2020

  5. [5]

    The mvtec 3d-addatasetforunsupervised3danomalydetectionandlocalization

    Bergmann, P., Jin, X., Sattlegger, D., Steger, C., 2021. The mvtec 3d-addatasetforunsupervised3danomalydetectionandlocalization. arXiv preprint arXiv:2112.09045

    arXiv 2021

  6. [6]

    Anomaly detection in 3d point clouds using deep geometric descriptors, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    Bergmann, P., 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, pp. 2613–2623

    work page 2023

  7. [7]

    3d-adcs: Untrained 3d anomaly detectionforcomplexmanufacturingsurfaces

    Cao, X., Tao, C., Du, J., 2025. 3d-adcs: Untrained 3d anomaly detectionforcomplexmanufacturingsurfaces. JournalofComputing and Information Science in Engineering , 1–15

    work page 2025

  8. [8]

    Informative knowledge distillation for image anomaly segmentation

    Cao, Y., Wan, Q., Shen, W., Gao, L., 2022. Informative knowledge distillation for image anomaly segmentation. Knowledge-Based Systems 248, 108846

    work page 2022

Show all 34 references
  1. [9]

    Complementarypseudomultimodal feature for point cloud anomaly detection

    Cao,Y.,Xu,X.,Shen,W.,2024. Complementarypseudomultimodal feature for point cloud anomaly detection. Pattern Recognition 156, 110761

    2024

  2. [10]

    Shapenet: An information-rich 3d model repository

    Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al., 2015. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012

    2015 arXiv

  3. [11]

    Shape- guided dual-memory learning for 3d anomaly detection, in: Proceed- ings of the 40th International Conference on Machine Learning, pp

    Chu, Y.M., Liu, C., Hsieh, T.I., Chen, H.T., Liu, T.L., 2023. Shape- guided dual-memory learning for 3d anomaly detection, in: Proceed- ings of the 40th International Conference on Machine Learning, pp. 6185–6194. Cao et al.:Preprint submitted to Elsevier Page 11 of 12

    2023

  4. [12]

    Padim:apatch distributionmodelingframeworkforanomalydetectionandlocaliza- tion,in:Internationalconferenceonpatternrecognition,Springer.pp

    Defard,T.,Setkov,A.,Loesch,A.,Audigier,R.,2021. Padim:apatch distributionmodelingframeworkforanomalydetectionandlocaliza- tion,in:Internationalconferenceonpatternrecognition,Springer.pp. 475–489

    2021

  5. [13]

    Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conferenceoncomputervisionandpatternrecognition,Ieee.pp.248– 255

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L., 2009. Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conferenceoncomputervisionandpatternrecognition,Ieee.pp.248– 255

    2009

  6. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy,A.,Beyer,L.,Kolesnikov,A.,Weissenborn,D.,Zhai,X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N., 2021. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR

    2021

  7. [15]

    3d vision-based anomaly detection in manufacturing: A survey

    Du, J., Tao, C., Cao, X., Tsung, F., 2025. 3d vision-based anomaly detection in manufacturing: A survey. Frontiers of Engineering Management , 1–18

    2025

  8. [16]

    Gudovskiy, D., Ishizaka, S., Kozuka, K., 2022. Cflow-ad: Real-time unsupervisedanomalydetectionwithlocalizationviaconditionalnor- malizing flows, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp. 98–107

    2022

  9. [17]

    Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778

    2016

  10. [18]

    Horwitz, E., 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, pp. 2968–2977

    2023

  11. [19]

    Randla-net: Efficient semantic segmentation of large-scalepointclouds,in:ProceedingsoftheIEEE/CVFconference on computer vision and pattern recognition, pp

    Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A., 2020. Randla-net: Efficient semantic segmentation of large-scalepointclouds,in:ProceedingsoftheIEEE/CVFconference on computer vision and pattern recognition, pp. 11108–11117

    2020

  12. [20]

    A multi-scale information fusion framework with interaction-aware global attention for industrial vi- sionanomalydetectionandlocalization.InformationFusion,103356

    Li, Z., Ge, Y., Meng, L., 2025. A multi-scale information fusion framework with interaction-aware global attention for industrial vi- sionanomalydetectionandlocalization.InformationFusion,103356

    2025

  13. [21]

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

    Lin, Y., Chang, Y., Tong, X., Yu, J., Liotta, A., Huang, G., Song, W., Zeng, D., Wu, Z., Wang, Y., et al., 2025. A survey on rgb, 3d, and multimodal approaches for unsupervised industrial image anomaly detection. Information Fusion , 103139

    2025

  14. [22]

    Memory association guided unsupervised anomaly detection with adaptive 3d attention

    Liu, X., Wu, C., Zhang, H., Wang, L., 2025. Memory association guided unsupervised anomaly detection with adaptive 3d attention. Information Fusion , 103379

    2025

  15. [23]

    Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces

    Ma, B., Han, Z., Liu, Y.S., Zwicker, M., 2020. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495

    2020 arXiv

  16. [24]

    Masked autoencoders for point cloud self-supervised learning, in: European conference on computer vision, Springer

    Pang, Y., Wang, W., Tay, F.E., Liu, W., Tian, Y., Yuan, L., 2022. Masked autoencoders for point cloud self-supervised learning, in: European conference on computer vision, Springer. pp. 604–621

    2022

  17. [25]

    Pointnet:Deeplearning on point sets for 3d classification and segmentation, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Qi,C.R.,Su,H.,Mo,K.,Guibas,L.J.,2017. Pointnet:Deeplearning on point sets for 3d classification and segmentation, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 652–660

    2017

  18. [26]

    Towards total recall in industrial anomaly detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., Gehler, P., 2022. Towards total recall in industrial anomaly detection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14318–14328

    2022

  19. [27]

    Asym- metric student-teacher networks for industrial anomaly detection, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp

    Rudolph, M., Wehrbein, T., Rosenhahn, B., Wandt, B., 2023. Asym- metric student-teacher networks for industrial anomaly detection, in: Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp. 2592–2602

    2023

  20. [28]

    Fast point feature his- tograms (fpfh) for 3d registration, in: 2009 IEEE international con- ference on robotics and automation, IEEE

    Rusu, R.B., Blodow, N., Beetz, M., 2009. Fast point feature his- tograms (fpfh) for 3d registration, in: 2009 IEEE international con- ference on robotics and automation, IEEE. pp. 3212–3217

    2009

  21. [29]

    G2sf-miad: Geometry-guided score fusion for multimodal industrial anomaly detection

    Tao, C., Cao, X., Du, J., 2025. G2sf-miad: Geometry-guided score fusion for multimodal industrial anomaly detection. arXiv preprint arXiv:2503.10091

    2025 arXiv

  22. [30]

    Knowl- edge fusion of large language models, in: The Twelfth International Conference on Learning Representations

    Wan, F., Huang, X., Cai, D., Quan, X., Bi, W., Shi, S., 2024. Knowl- edge fusion of large language models, in: The Twelfth International Conference on Learning Representations

    2024

  23. [31]

    Multimodal industrial anomaly detection via hybrid fusion, in: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Wang, Y., Peng, J., Zhang, J., Yi, R., Wang, Y., Wang, C., 2023. Multimodal industrial anomaly detection via hybrid fusion, in: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8032–8041

    2023

  24. [32]

    Invase: Instance-wise variableselectionusingneuralnetworks,in:Internationalconference on learning representations

    Yoon, J., Jordon, J., Van der Schaar, M., 2018. Invase: Instance-wise variableselectionusingneuralnetworks,in:Internationalconference on learning representations

    2018

  25. [33]

    Destseg:Segmentationguideddenoisingstudent-teacherforanomaly detection,in:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition, pp

    Zhang, X., Li, S., Li, X., Huang, P., Shan, J., Chen, T., 2023. Destseg:Segmentationguideddenoisingstudent-teacherforanomaly detection,in:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition, pp. 3914–3923. Cao et al.:Preprint submitted to Elsevier Page 12 of 12

    2023

  26. [2025]

    Information Fusion , 103517

    FoundationmodelsandTransformersforanomalydetection:A survey. Information Fusion , 103517

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.