Pith. sign in

REVIEW 2 major objections 5 minor 38 references

ProCon: Projection-Consistency Memory for Training-Free Anomaly Detection

T0 review · 2 major / 5 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Soft projection of test patches onto normal memory neighborhoods turns hard nearest-neighbor scoring into decoder-free reconstruction and improves industrial anomaly localization.

desk verdict Solid training-free redesign of PatchCore-style scoring: soft local projection residual plus residual-aligned consensus, with clean ablations and public code. read the letter →

arxiv 2607.04894 v1 pith:DQOZ3Q6W submitted 2026-07-06 cs.CV

Joongwon Chae , Lihui Luo , Yang Liu , Dongmei Yu , Peiwu Qin , Runming Wang , Ilmoon Chae This is my paper · ORCID
classification cs.CV
keywords anomalydetectionmemorybanksoftprojectionresidualtraining-freeDINOv2industrialinspectionconsensus
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

Industrial anomaly detectors that store only normal patch features usually decide normality by asking whether a test patch has one nearby normal anchor. That hard nearest-neighbor test can accept a defective patch that happens to match a single stored vector even when the local normal neighborhood cannot explain it. ProCon replaces the hard lookup with a soft projection: each test patch is reconstructed as a distance-weighted combination of its nearest normal memory vectors, and the residual of that projection becomes the anomaly score. The residual is stabilized by two consensus steps that never train anything—median aggregation over seed-perturbed coreset banks and mean aggregation over independent residual maps from four frozen transformer depths. On the standard single-category benchmarks the resulting maps and image scores match or exceed recent reconstruction and Transformer methods while remaining completely training-free. The practical claim is that memory should be treated as a local normal support set queried for explainability rather than as a table of exemplars.

What carries the argument

Soft local normal projection: for each test patch the k nearest normal memory vectors receive distance-based soft weights, their weighted combination is the projected normal estimate, and the Euclidean residual of that projection is the anomaly evidence; bank-wise median and layer-wise mean consensus then stabilize the residual without any learned fusion.

What would settle it

A controlled experiment in which a modest fraction of repeated abnormal patterns is deliberately added to the normal training set (or the four selected depths are replaced by shallow texture-only layers) and the pixel-level residual maps on texture-dominated categories such as toy bricks or mints cease to rank defects above background.

Watch

Extended reading notes

Core claim

Replacing hard nearest-neighbor memory scoring with soft local normal projection residuals, then stabilizing those residuals by bank-wise median and layer-wise mean consensus over independent DINOv2 layer memories, yields stronger image- and pixel-level anomaly detection than hard retrieval and competitive or better results than recent reconstruction and Transformer methods on MVTec-AD, VisA and Real-IAD under the single-category protocol.

Load-bearing premise

That a frozen vision-transformer feature space, a fixed four-layer pool, and a pure normal coreset together supply a local neighborhood whose soft-projection residual reliably flags defects for every industrial category.

Editorial extensions

If this is right

  • Memory banks can serve as non-parametric reconstruction operators, removing the need to train a decoder or synthesize pseudo anomalies while still producing residual-style scores.
  • False-normal matches that plague nearest-neighbor scoring are reduced once a patch must be explained by a whole local neighborhood rather than a single anchor.
  • Independent per-layer memories fused after residual alignment preserve depth-specific normal geometry that is lost when features are concatenated before coreset selection.
  • The same fixed projection-and-consensus recipe transfers across consumer-object, metal-part, and medical benchmarks without retuning, indicating that the readout itself, not dataset-specific calibration, drives the gains.

Reading between the lines

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

  • As stronger frozen encoders appear, the same soft-projection residual can be re-used without redesigning a decoder, making the method a lightweight way to harvest encoder improvements for anomaly localization.
  • Approximate nearest-neighbor search or compressed coresets would preserve the scoring objective while lowering the 20 layer-bank retrievals per image, making the approach practical for high-throughput inspection lines.
  • Texture-dominated categories that remain hard under the current residual suggest that an explicit multi-scale texture memory branch, still training-free, could close the remaining gap without leaving the projection-consistency framework.
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

2 major / 5 minor

Summary. ProCon reinterprets PatchCore-style nearest-neighbor memory scoring as hard projection onto a single normal anchor and replaces it with soft local normal projection: each test patch is reconstructed as a distance-weighted combination of its k nearest normal memory vectors, and the residual is used as decoder-free anomaly evidence. Residuals are stabilized by two consensus axes—median over B seed-perturbed coreset banks per layer, and mean over independent residual maps from a fixed DINOv2 layer pool {−3,−6,−8,−9}—with fusion performed only after residual alignment. The method is fully training-free (frozen backbone, no decoder, no learned fusion, no pseudo-anomalies). On MVTec-AD, VisA, and Real-IAD under the single-category protocol it reports image AUROC 99.8/99.2/93.2 and competitive or superior pixel AP and AUPRO versus recent memory, reconstruction, and Transformer baselines. Controlled genealogy, independent-vs-concat, and coreset-budget ablations attribute the gains to the projection residual and residual-aligned consensus rather than memory size.

Significance. If the empirical claim holds, ProCon supplies a clean, reproducible middle ground between memory retrieval and learned reconstruction: reconstruction-style residuals without decoder training or pseudo-anomaly synthesis, obtained by a fixed non-parametric operator on frozen DINOv2 features. The controlled ablation ladder (hard NN → bank consensus → soft projection → layer consensus), the independent-vs-concat comparison, and the budget result that ProCon@1% already beats the single-layer soft-projection baseline@10% make the source of the gain unusually transparent for this literature. Public code, a fully specified recipe (including adaptive temperature and fixed random projection), seven-metric tables, and cross-domain transfer to MPDD/BTAD/Uni-Medical further raise the bar for training-free industrial AD. The contribution is incremental rather than paradigm-shifting, but it is carefully executed and immediately usable.

major comments (2)
  1. Main comparison (Table 1) mixes published baseline numbers with ProCon under a DINOv2 ViT-B/14 backbone and a 5%/1% coreset recipe. While the controlled SPM-vs-ProCon tables (Tables 2, 6, 8) isolate the readout change under a fixed backbone, the headline claim of superiority over Dinomaly/INP-Former would be stronger if at least one strong reconstruction baseline were re-run under the same backbone and evaluation seed, or if the authors explicitly quantified the backbone contribution. As written, backbone and readout are partially confounded in the external comparison.
  2. Complexity and throughput are stated only asymptotically (O(|L| B P |M| D) with 20 layer-bank evaluations per image). For a training-free industrial method this is load-bearing: without wall-clock latency or memory-footprint numbers against PatchCore/SPM on the same hardware, it is hard to judge whether the residual consensus is practical for high-throughput inspection, a limitation the authors themselves flag but do not quantify.
minor comments (5)
  1. Notation for the adaptive temperature τ_ℓ,b is clear in the method, but the main text does not state that the median is computed on a 512-patch subsample; this efficiency detail appears only in the supplement and should be mentioned once in the main method for reproducibility.
  2. Figure 1 is schematic only; a small quantitative panel (e.g., residual histograms for normal vs anomalous patches under hard NN vs soft projection) would make the false-normal-match argument more concrete without expanding the page budget.
  3. Self-citations to concurrent StructCore and GCR arXiv preprints are appropriate for context but should be clearly marked as concurrent/related work so that readers do not treat them as established prior art.
  4. Typographical spacing issues appear throughout (e.g., “localizesdefects”, “isconsistentlysupported”); a careful pass would improve readability.
  5. The fixed random projection (768→512, plus 192-d sub-projection for coreset distances) is fully specified in the supplement; a one-sentence pointer in the main Method section would help readers who do not immediately open the appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical training-free method with controlled ablations on external public benchmarks; self-citations are non-load-bearing related-work context.

full rationale

ProCon is an empirical computer-vision method paper. Its central claim—that soft local normal projection residuals (with bank-wise median and residual-aligned layer-wise mean consensus over independent DINOv2 coreset memories) improve training-free memory anomaly detection over hard nearest-neighbor retrieval—is supported by a controlled genealogy ladder (Tables 2/8), independent-vs-concat comparison (Tables 3/9), coreset-budget sweeps showing ProCon@1% already beats the single-layer SPM baseline@10% (Tables 4/16–17), seven-metric results on MVTec-AD/VisA/Real-IAD, and cross-domain transfer (Tables 22–25). Temperature is set adaptively from each test image’s own local distance median; coreset ratios and the fixed layer pool are chosen by saturation/promotion analysis, not fitted to the reported target metrics. The concurrent self-citations (StructCore, GCR) appear only in Related Work as memory-redesign context and are never invoked as uniqueness theorems, forced ansätze, or premises that make the ProCon residual claim true by construction. No equation reduces to its own inputs; no prediction is statistically forced by a fit; the evaluation is against external public benchmarks under a standard single-category protocol. Therefore the derivation chain is self-contained and non-circular.

Assumptions & free parameters 6 free parameters · 3 assumptions · 2 invented entities

The method rests on standard pretrained-feature assumptions plus a handful of fixed design choices (layer pool, k, B, coreset ratios, adaptive temperature). No new physical entities are postulated; the free parameters are ordinary hyper-parameters of a training-free pipeline.

free parameters (6)
  • layer pool L = {-3,-6,-8,-9}
    Selected by multi-phase validation and promotion; fixed across all final experiments and not derived from first principles.
  • number of seed banks B=5
    Chosen after bank-count sweep; gains saturate early but B remains a free design choice.
  • soft-projection neighborhood size k=5
    Fixed hyper-parameter controlling the local normal support set.
  • coreset ratio ρm (5% MVTec/VisA, 1% Real-IAD)
    Budget chosen from saturation analysis; different values per dataset family.
  • top-mean image readout fraction ρ=0.005
    Fixed pooling ratio for converting maps to image scores.
  • adaptive temperature τ = median local squared distances
    Data-dependent but rule-based; still a design choice that sets softmax bandwidth.
assumptions (3)
  • domain assumption Frozen DINOv2 ViT-B/14 patch features form a geometry in which local normal neighborhoods can reconstruct normal patches better than anomalous ones.
    Invoked throughout Method and Discussion; standard in modern memory-based AD but not proved.
  • domain assumption Greedy farthest-point coreset at small ratios preserves sufficient normal support for soft projection.
    Used in memory construction; inherited from PatchCore-style practice.
  • ad hoc to paper Median over seed-perturbed banks and mean over residual-aligned layers are valid consensus operators that reduce false-normal matches.
    Central design claim of ProCon; supported by ablations but not derived from a uniqueness theorem.
invented entities (2)
  • soft local normal projection residual as decoder-free anomaly evidence independent evidence
    purpose: Replace hard nearest-neighbor distance with a reconstruction-style residual computed non-parametrically from memory.
    Core technical object of the paper; independent evidence is the empirical localization gains under controlled budgets.
  • projection-consistency memory (ProCon) with bank- and layer-wise residual consensus independent evidence
    purpose: Stabilize the residual by enforcing residual alignment before fusion.
    Named framework that packages the soft residual with two consensus axes; evidence is the ablation ladder.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ProCon: Projection-Consistency Memory for Training-Free Anomaly Detection." pith.science (2026). https://pith.science/paper/DQOZ3Q6W

@misc{pith2026260704894,
  author       = {Pith},
  title        = {Pith review of: ProCon: Projection-Consistency Memory for Training-Free Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DQOZ3Q6W}},
  note         = {Machine review of arXiv:2607.04894}
}
read the original abstract

Memory-based anomaly detection is attractive because it localizes defects from normal images without training a decoder or synthesizing pseudo anomalies. However, most memory methods still use the memory bank as a nearest-neighbor lookup table: a test patch is treated as normal if it has one nearby normal anchor. This hard retrieval view is vulnerable to false-normal matches and does not test whether the patch is consistently supported by a local normal neighborhood. We propose ProCon, a training-free framework that turns memory retrieval into decoder-free reconstruction. ProCon softly projects each test patch onto nearby normal memory vectors and uses the projection residual as anomaly evidence. To stabilize this residual, it constructs seed-perturbed layer-wise memories, aggregates bank residuals by a median, and fuses depth-specific residual maps by layer consensus. ProCon requires no decoder training, backbone fine-tuning, learned fusion weights, or pseudo-anomaly supervision. Across MVTec-AD, VisA, and Real-IAD under the single-category evaluation protocol, ProCon achieves strong image- and pixel-level performance under seven standard metrics, including image AUROC scores of 99.8%, 99.2%, and 93.2%, respectively. Ablations show that the gains come from replacing hard retrieval with soft normal projection and stabilizing the residuals through memory and depth consensus. The code is available at https://github.com/jw-chae/Procon

Figures

Figures reproduced from arXiv: 2607.04894 by the authors.

Figure 1
Figure 1. Overview of ProCon. Normal images are encoded by a frozen DINOv2 backbone, and independent layer-wise coreset [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Compact qualitative examples. ProCon localizes [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Layer-wise residual maps produced by ProCon. Each selected depth generates a residual map under the same soft [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Additional qualitative examples on MVTec-AD, part 1. Rows show bottle, cable, capsule, carpet, and grid. [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Additional qualitative examples on MVTec-AD, part 2. Rows show hazelnut, leather, metal nut, pill, and screw. [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Additional qualitative examples on MVTec-AD, part 3. Rows show tile, toothbrush, transistor, wood, and zipper. [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Additional qualitative examples on VisA, part 1. Rows show candle, capsules, cashew, chewing gum, and fryum. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]
Figure 8
Figure 8. Figure 8: Additional qualitative examples on VisA, part 2. Rows show macaroni1, macaroni2, pcb1, pcb2, and pcb3. [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Additional qualitative examples on VisA, part 3. Rows show pcb4 and pipe fryum. [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 10
Figure 10. Figure 10: Additional qualitative examples on Real-IAD, part 1. Rows show audiojack, bottle cap, button battery, end cap, and [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: Additional qualitative examples on Real-IAD, part 2. Rows show fire hood, mint, mounts, pcb, and phone battery. [PITH_FULL_IMAGE:figures/full_fig_p027_11.png]
Figure 12
Figure 12. Figure 12: Additional qualitative examples on Real-IAD, part 3. Rows show plastic nut, plastic plug, porcelain doll, regulator, [PITH_FULL_IMAGE:figures/full_fig_p028_12.png]
Figure 13
Figure 13. Figure 13: Additional qualitative examples on Real-IAD, part 4. Rows show sim card set, switch, tape, terminalblock, and [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]
Figure 14
Figure 14. Figure 14: Additional qualitative examples on Real-IAD, part 5. Rows show toy, toy brick, transistor1, u block, and usb. [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]
Figure 15
Figure 15. Figure 15: Additional qualitative examples on Real-IAD, part 6. Rows show usb adaptor, vcpill, wooden beads, woodstick, and [PITH_FULL_IMAGE:figures/full_fig_p031_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 1 canonical work pages

  1. [1]

    Proceedings of the

    Batzner, Kilian and Heckler, Lars and K. Proceedings of the. 2024 , url =

    2024

  2. [2]

    Bergmann, Paul and Fauser, Michael and Sattlegger, David and Steger, Carsten , year =. 2019. doi:10.1109/CVPR.2019.00982 , url =

    doi:10.1109/cvpr.2019.00982 2019

  3. [3]

    1996 , journal =

    Bagging Predictors , author =. 1996 , journal =. doi:10.1007/BF00058655 , url =

    doi:10.1007/bf00058655 1996

  4. [4]

    2601.01856 , archivePrefix =

    Chae, Joongwon and Luo, Lihui and Liu, Yang and Wang, Runming and Yu, Dongmei and Liang, Zeming and Yuan, Xi and Zhang, Dayan and Chen, Zhenglin and Qin, Peiwu and Chae, Ilmoon , year =. 2601.01856 , archivePrefix =

    arXiv

  5. [5]

    2602.17048 , archivePrefix =

    Chae, Joongwon and Luo, Lihui and Liu, Yang and Wang, Runming and Yu, Dongmei and Liang, Zeming and Yuan, Xi and Zhang, Dayan and Chen, Zhenglin and Qin, Peiwu and Chae, Ilmoon , year =. 2602.17048 , archivePrefix =

    arXiv

  6. [6]

    2020 , eprint =

    Sub-Image Anomaly Detection with Deep Pyramid Correspondences , author =. 2020 , eprint =

    2020

  7. [7]

    Pattern Recognition

    Defard, Thomas and Setkov, Aleksandr and Loesch, Angelique and Audigier, Romaric , year =. Pattern Recognition

  8. [8]

    Anomaly Detection via Reverse Distillation from One-Class Embedding , booktitle =

    Deng, Hanqiu and Li, Xingyu , year =. Anomaly Detection via Reverse Distillation from One-Class Embedding , booktitle =

Show all 38 references
  1. [9]

    , year =

    Dietterich, Thomas G. , year =. Ensemble Methods in Machine Learning , booktitle =. doi:10.1007/3-540-45014-9_1 , url =

    doi

  2. [10]

    An Image Is Worth 16x16 Words:

    Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil , year =. An Image Is Worth ...

  3. [11]

    Proceedings of the

    Gudovskiy, Denis and Ishizaka, Shun and Kozuka, Kazuki , year =. Proceedings of the

  4. [12]

    Dinomaly:

    Guo, Jia and Lu, Shuai and Zhang, Weihang and Chen, Fang and Li, Huiqi and Liao, Hongen , year =. Dinomaly:. Proceedings of the

  5. [13]

    1990 , journal =

    Neural Network Ensembles , author =. 1990 , journal =. doi:10.1109/34.58871 , url =

    1990 doi

  6. [14]

    1998 , journal =

    The Random Subspace Method for Constructing Decision Forests , author =. 1998 , journal =. doi:10.1109/34.709601 , url =

    1998 doi

  7. [15]

    Divide-and-Assemble:

    Hou, Jinlei and Zhang, Yingying and Zhong, Qiaoyong and Xie, Di and Pu, Shiliang and Zhou, Hong , year =. Divide-and-Assemble:. Proceedings of the

  8. [16]

    Proceedings of the

    Hyun, Jeeho and Kim, Sangyun and Jeon, Giyoung and Kim, Seung Hwan and Bae, Kyunghoon and Kang, Byung Jun , year =. Proceedings of the

  9. [17]

    Advances in Neural Information Processing Systems (

    Jiang, Xi and Liu, Jianlin and Wang, Jinbao and Nie, Qiang and Wu, Kai and Liu, Yong and Wang, Chengjie and Zheng, Feng , year =. Advances in Neural Information Processing Systems (

  10. [18]

    Neural Network Ensembles, Cross Validation, and Active Learning , booktitle =

    Krogh, Anders and Vedelsby, Jesper , year =. Neural Network Ensembles, Cross Validation, and Active Learning , booktitle =

  11. [19]

    Feature Bagging for Outlier Detection , booktitle =

    Lazarevic, Aleksandar and Kumar, Vipin , year =. Feature Bagging for Outlier Detection , booktitle =. doi:10.1145/1081870.1081891 , url =

    doi

  12. [20]

    IEEE access : practical innovations, open solutions , volume =

    Lee, Sungwook and Lee, Seunghyun and Song, Byung Cheol , year =. IEEE access : practical innovations, open solutions , volume =. doi:10.1109/ACCESS.2022.3193699 , url =

    2022 doi

  13. [21]

    Proceedings of the

    Liu, Zhikang and Zhou, Yiming and Xu, Yuansheng and Wang, Zilei , year =. Proceedings of the

  14. [22]

    Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection , booktitle =

    Luo, Wei and Cao, Yunkang and Yao, Haiming and Zhang, Xiaotian and Lou, Jianan and Cheng, Yuqi and Shen, Weiming and Yu, Wenyong , year =. Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection , booktitle =

  15. [23]

    2023 , eprint =

    Oquab, Maxime and Darcet, Timoth. 2023 , eprint =

    2023

  16. [24]

    Less Is More:

    Rayana, Shebuti and Akoglu, Leman , year =. Less Is More:. ACM Transactions on Knowledge Discovery from Data , volume =

  17. [25]

    Towards Total Recall in Industrial Anomaly Detection , booktitle =

    Roth, Karsten and Pemula, Latha and Zepeda, Joaquin and Sch. Towards Total Recall in Industrial Anomaly Detection , booktitle =. 2022 , pages =

    2022

  18. [26]

    1996 , journal =

    Error Correlation and Error Reduction in Ensemble Classifiers , author =. 1996 , journal =. doi:10.1080/095400996116839 , url =

    1996 doi

  19. [27]

    Wang, Chengjie and Zhu, Wenbing and Gao, Bin-Bin and Gan, Zhenye and Zhang, Jiangning and Gu, Zhihao and Qian, Shuguang and Chen, Mingang and Ma, Lizhuang , year =. Real-. Proceedings of the

  20. [28]

    Pushing the Limits of Fewshot Anomaly Detection in Industry Vision:

    Xie, Guoyang and Wang, Jinbao and Liu, Jiaqi and Zheng, Feng and Jin, Yaochu , year =. Pushing the Limits of Fewshot Anomaly Detection in Industry Vision:. International Conference on Learning Representations (

  21. [29]

    A Unified Model for Multi-Class Anomaly Detection , booktitle =

    You, Zhiyuan and Cui, Lei and Shen, Yujun and Yang, Kai and Lu, Xin and Zheng, Yu and Le, Xinyi , year =. A Unified Model for Multi-Class Anomaly Detection , booktitle =

  22. [30]

    2111.07677 , archivePrefix =

    Yu, Jiawei and Zheng, Ye and Wang, Xiang and Li, Wei and Wu, Yushuang and Zhao, Rui and Wu, Liwei , year =. 2111.07677 , archivePrefix =

    arXiv

  23. [31]

    Proceedings of the

    Zavrtanik, Vitjan and Kristan, Matej and Sko. Proceedings of the. 2021 , pages =

    2021

  24. [32]

    Proceedings of the

    Zhang, Xuan and Li, Shiyu and Li, Xi and Huang, Ping and Shan, Jiulong and Chen, Ting , year =. Proceedings of the

  25. [33]

    and Li, Zheng , year =

    Zhao, Yue and Nasrullah, Zain and Hryniewicki, Maciej K. and Li, Zheng , year =. Proceedings of the 2019. doi:10.1137/1.9781611975673.66 , url =

    2019 doi

  26. [34]

    Zimek, Arthur and Campello, Ricardo J. G. B. and Sander, J. Ensembles for Unsupervised Outlier Detection:. 2013 , journal =. doi:10.1145/2594473.2594476 , url =

    2013 doi

  27. [35]

    Proceedings of the European Conference on Computer Vision (

    Zou, Yang and Jeong, Jongheon and Pemula, Latha and Zhang, Dongqing and Dabeer, Onkar , year =. Proceedings of the European Conference on Computer Vision (

  28. [36]

    Deep Learning-Based Defect Detection of Metal Parts: Evaluating Current Methods in Complex Conditions , booktitle =

    Jezek, Stepan and Jonak, Martin and Burget, Radim and Dvorak, Pavel and Skotak, Milos , year =. Deep Learning-Based Defect Detection of Metal Parts: Evaluating Current Methods in Complex Conditions , booktitle =

  29. [37]

    2021 IEEE 30th International Symposium on Industrial Electronics (

    Mishra, Pankaj and Verk, Riccardo and Fornasier, Daniele and Piciarelli, Claudio and Foresti, Gian Luca , year =. 2021 IEEE 30th International Symposium on Industrial Electronics (

    2021

  30. [38]

    Proceedings of the

    Bao, Jinan and Sun, Hanshi and Deng, Hanqiu and He, Yinsheng and Zhang, Zhaoxiang and Li, Xingyu , year =. Proceedings of the

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.