Pith. sign in

REVIEW 4 major objections 5 minor 35 references

HomographyAD: Deep Anomaly Detection Using Self Homography Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Input foreground alignment via deep homography estimation, followed by self-homography fine-tuning of an ImageNet-pretrained backbone, improves pretrained anomaly detectors on object-centric industrial datasets.

desk verdict Test-set tuning of the fine-tuning iteration invalidates the paper's main claim, though the self-homography regression idea has some novelty. read the letter →

arxiv 2506.08784 v1 pith:KNXFG2AC submitted 2025-06-10 cs.CV

Jongyub Seok , Chanjin Kang This is my paper
classification cs.CV
keywords anomalydetectiondeephomographyestimationself-supervisedlearningpretrainednetworksindustrialinspectionimagealignmentMVTecshapeinformation
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

HomographyAD is built on the observation that pretrained anomaly detectors assume the same image location always shows the same object part, while real industrial images violate this because the foreground is not aligned. The paper shows that degrading alignment on MVTec data lowers image-level AUROC of PaDiM, PatchCore, SPADE, and MahAD, and that restoring alignment with deep homography estimation recovers and sometimes improves performance. It then fine-tunes the ImageNet-pretrained backbone with a regression proxy task, predicting the four-corner displacement of an applied 2D transform on normal samples, so the network learns shape information that texture-biased ImageNet features lack. Applied to several PAD methods, this self-homography learning raises image-level AUROC on object classes while slightly lowering it on texture classes, a trade-off the authors attribute to losing texture information and leave as future work.

What carries the argument

The central machinery is deep homography estimation plus a regression-based self-supervised task. A homography is the 2D projective transformation mapping one planar image to another; the deep homography model predicts the eight corner displacements $(\Delta x_i, \Delta y_i)$ and converts them into a homography matrix by DLT, aligning every input to a template image so that position-consistent feature comparisons are valid. Self-homography learning applies a random 2D perturbation to each aligned normal image and asks the ImageNet-pretrained backbone (with a modified head) to regress those displacements under the loss $L = \sum_{i=1}^{4} |\Delta_i - \Delta'_i|^2$, injecting shape information into features that are otherwise texture-biased. The final anomaly score is the distance of test features from the distribution of normal features, following the PAD method being wrapped.

What would settle it

Re-run the Section 4.3.2 comparison choosing the fine-tuning checkpoint by validation-split AUROC or a fixed schedule instead of the highest test AUROC per class; if object-class gains over the ImageNet baseline do not survive, the reported benefit rests on test-set selection rather than on homography learning.

Watch

Extended reading notes

Core claim

The paper's central claim is that input foreground alignment is a key factor for pretrained anomaly detection in industrial settings, and that two uses of homography, aligning inputs before feature extraction and fine-tuning the backbone with self-homography learning, can turn a poorly performing PAD method into a better one for object-centric classes. On the synthetic alignment experiments, every tested PAD method's image-level AUROC drops when MVTec images are deliberately misaligned, and most recover or improve when images are aligned via a learned homography. The self-homography stage regresses random four-corner perturbations on aligned normal images, using an L2 loss on the eight displacement values, which forces the feature extractor to attend to shape rather than texture; across PaDiM, PatchCore, SPADE, and MahAD, and across ResNet18, WideResNet50, and EfficientNet-B5 backbones, object-class AUROC increases. The authors note the reverse holds for texture classes, where the fine-tuned network loses some performance, and propose preserving texture information and an end-to-end design as future work.

Load-bearing premise

The load-bearing premise, stated in Section 4.2, is that selecting the fine-tuning iteration with the highest average test AUROC per class is a valid way to measure the method's gain; if that selection is an artifact of test-set peeking, the claimed benefit is not established.

Editorial extensions

If this is right

  • Input-level foreground alignment is a drop-in preprocessing step: any PAD method whose features assume position consistency should gain from deep homography alignment before feature extraction.
  • Self-homography learning provides a label-free fine-tuning recipe that transfers across PAD methods such as PaDiM, PatchCore, SPADE, and MahAD, and across backbones such as ResNet18, WideResNet50, and EfficientNet-B5.
  • Object-centric industrial classes are where the shape-information boost shows up; texture classes need augmentation tuning to avoid a small regression in performance.
  • Because the method uses regression to continuous corner displacements rather than classification of discrete rotations, it can capture a continuous range of 2D transformations in the learned representation.

Reading between the lines

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

  • The best-test-iteration selection in Section 4.2 means the practical gain depends on having a validation protocol; a fixed schedule or validation-based early stopping could reduce the reported object-class improvements.
  • The same two-stage recipe could be tested on pose-varying inspection tasks beyond MVTec, such as electronics assembly or automotive parts, where natural misalignment is the norm; an end-to-end version would also remove the cost of training a separate alignment model.
  • A hybrid self-supervised objective that adds a texture-preserving term, such as reconstruction or contrastive loss, to the homography regression could counteract the reported texture-class drop while keeping shape gains.
  • The alignment-plus-self-supervised-learning recipe transfers naturally to video frames of moving objects, where the continuous regression supervision could double as a pose-tracking signal.
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 / 5 minor

Summary. The paper proposes HomographyAD, a framework for unsupervised anomaly detection in industrial images, with two main components: input-level foreground alignment using deep homography estimation, and self-homography learning, in which an ImageNet-pretrained backbone is fine-tuned to regress random 2D corner perturbations of normal images. Anomaly detection is then performed with existing pretrained-anomaly-detection (PAD) methods, namely PaDiM, PatchCore, SPADE, and MahAD, on top of the fine-tuned features. The authors evaluate on the MVTec dataset and on synthetic MVTec variants with different alignment conditions, reporting that alignment improves PAD performance and that self-homography learning improves object-class performance while degrading texture-class performance.

Significance. The problem addressed is real and timely: pretrained-feature anomaly detection methods assume spatial alignment, which is often violated in industrial settings. The self-homography regression task is simple and could in principle be applied to any PAD method, and the paper includes experiments across multiple PAD methods, backbones, and augmentation settings. However, the central empirical claim of performance enhancement is not established because the main experiments select the fine-tuning checkpoint using test-set AUROC, and the reported texture-class degradations are large. If the experimental protocol were corrected and the claims appropriately qualified, the idea could be a useful contribution, but as submitted the evidence does not support the abstract's claim of general improvement.

major comments (4)
  1. [Section 4.2 (Table 2)] The protocol for reporting PAD w/ HL results is test-set-optimized. The text states that fine-tuning is performed for 3000 iterations, AUROC is measured every 100 iterations, and 'the iteration with the highest average AUROC was selected as the representative value.' This means every PAD w/ HL number in Table 2 is a maximum over 30 checkpoints chosen using test-set labels, whereas the PAD baseline is a single fixed pretrained network. The reported gains, such as +2.94 for PaDiM object-level and +1.52 for MahAD image-level, are therefore upper bounds selected with test information and do not represent the performance of any fixed training schedule. This invalidates the comparison as evidence that self-homography learning improves PAD in a predictive sense. The authors need to report results with a predetermined iteration count, or select the iteration on a validation split, and report the variation across iterations and runs.
  2. [Section 4.3.2 (Table 2)] The method causes severe degradations on texture classes that are not reflected in the abstract's general claim. For example, SPADE image-level AUROC drops from 95.32 to 62.67 (-32.66), and MahAD image-level AUROC drops from 92.02 to 68.80 (-23.21). The total image-level AUROC for SPADE and MahAD decreases by 7.78 and 7.12, respectively. Since these are already the best checkpoints selected on test AUROC, the actual degradation under a fixed schedule could be even larger. The paper's claim of 'performance enhancement through extensive experiments' is therefore not supported for these methods, and the object/texture trade-off must be addressed or the claims substantially qualified.
  3. [Section 4.1.2 (Table 1)] The alignment hypothesis is tested only on synthetic MVTec variants of four object classes (screw, hazelnut, metalnut, grid) that are generated using the authors' own deep homography model and rotation-only transformations. This does not directly validate the method on real-world misaligned industrial data, which is the motivating scenario. In addition, the PatchCore drop on Synthetic MVTec w/ Alignment is attributed to padding artifacts, but no experiment is provided to verify that explanation. A real-world misaligned dataset or a more controlled synthetic benchmark with varied transformation types is needed to support the first contribution.
  4. [Section 4.2 (Tables 2-4)] All reported AUROC values appear to come from a single run with no error bars, confidence intervals, or significance tests. Given the small differences in several comparisons, such as the +0.05 total image-level AUROC for EfficientNet B5 in Table 4, and the checkpoint-selection protocol, the reader cannot determine whether any of the improvements are reliable or are artifacts of run-to-run variation. Multiple runs with random seeds, or at least a sensitivity analysis over iterations, are needed for the central claims.
minor comments (5)
  1. [Section 5.1 (Table 3)] The column labeled 'Max' in Table 3 is not defined in the text; it appears to be the maximum over augmentation combinations, but this should be stated explicitly and the selection criterion should be given.
  2. [Section 3.1] The paper says augmentation includes 'pepper and salt' and later calls it 'shape augmentation' in Section 5.1, but pepper-and-salt noise is a noise augmentation, not a shape transform; this terminology is confusing and should be clarified.
  3. [Throughout] There are numerous typos and formatting issues, such as 'anomal part', 'Commonly', inconsistent 'MVtec' vs 'MVTec', and the truncated reference formatting for 'Schleget al.' These should be corrected in a revision.
  4. [Section 3.1] The choice of the template image and the procedure for applying the inferred homography at test time are not described in enough detail to be reproduced; the paper should specify how the template is selected and whether alignment is applied per class or globally.
  5. [Section 5.2] The claim that self-homography learning has 'the same effect' across backbones is only shown for PaDiM; the paper should either test other PAD methods on multiple backbones or restrict the claim to PaDiM.

Circularity Check

2 steps flagged · score 6.0 of 10

Reported gains are test-optimized: Section 4.2 picks the fine-tuning iteration with the highest test AUROC, so Table 2's 'PAD w/ HL' numbers are maxima over test-labeled checkpoints rather than predictions.

  1. fitted input called prediction [Section 4.2 (Experimental Setup) and Table 2]
    "Fine tuning performed 3000 iterations by each class and measured AUROC by each 100 iteration. Among the performances of each iteration, the iteration with the highest average AUROC was selected as the representative value."

    The reported PAD w/ HL AUROC in Table 2 is not the performance of a fixed training schedule; it is the maximum over 30 checkpoints chosen using test-set AUROC. The gains over the ImageNet-pretrained baseline are therefore test-optimized values, not predictions of an independently trained model. By construction, the headline 'performance enhancement' reduces to selecting the checkpoint most favorable on the evaluation set, so the central claim in the abstract is not established as a generalization result.

  2. fitted input called prediction [Section 5.1 / Table 3 (augmentation analysis)]
    "The final AD performance using self homography fine tuned network was different depending on which combination of augmentation was applied. ... PAD w/ HL Max"

    Table 3 reports a 'Max' column across augmentation combinations per class, again chosen using the evaluation AUROC. The claim that self homography learning can improve PAD is thus supported by the existence of a test-favorable augmentation configuration rather than by a fixed-procedure improvement; the same test-driven selection inflates the reported comparison and makes the conclusion partially circular with respect to the evaluation metric.

full rationale

The paper's pipeline is empirical: deep homography alignment, self-supervised homography fine-tuning with the L1-style regression loss in Eq. (1), and existing PAD feature-scoring. The loss target is a random 2D perturbation independent of anomaly labels, and the PAD backbones are externally pretrained, so there is no definitional circularity in the method itself and no load-bearing self-citation chain. However, the central evidence for the claim that homography learning improves PAD is Table 2, and its protocol in Section 4.2 selects the fine-tuning iteration with the highest test AUROC over 3000 iterations. Each reported 'PAD w/ HL' number is therefore a maximum over test-evaluated checkpoints, not a prediction from a fixed training schedule. The same test-driven selection appears in Table 3's 'Max' augmentation column. This makes the claimed 'performance enhancement' partially reduce by construction to choosing the most test-favorable configuration, though the underlying method and loss remain independent of AD labels. Score 6 reflects partial circularity in the evaluation of the central claim, not full definitional circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claims rest on several qualitative domain assumptions about texture bias, alignment benefit, and DHE generalization, plus two free hyperparameters. No new entities are introduced. The most damaging element is the free selection of fine-tuning iterations based on test AUROC, which appears as a fitted parameter.

free parameters (2)
  • Fine-tuning iterations (selected on test AUROC) = 3000 iterations, best AUROC chosen per class
    Section 4.2 states the iteration with the highest average AUROC is selected as the representative value. This is a free parameter tuned to the test set, which can inflate performance.
  • Perturbation range for self-homography learning = Not specified (random perturbation directed inside image)
    The paper does not report the distribution or amplitude of corner perturbations, a key hyperparameter for the SSL task.
assumptions (3)
  • domain assumption ImageNet pretrained features are texture-biased and lack shape information, so a fine-tuning task that enforces shape awareness will improve AD on object classes.
    Motivates the self-homography task (Section 3.2), borrowing from Geirhos et al. but not empirically tested for AD in this paper.
  • domain assumption Better alignment of object foreground in input images improves PAD performance.
    Stated as an assumption in Section 1 and only tested on synthetic MVTec variants generated with the authors' own DHE model.
  • domain assumption The DHE model trained on normal samples generalizes to test samples of the same class.
    Used in Section 3.1 for input alignment; no evaluation of alignment accuracy is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HomographyAD: Deep Anomaly Detection Using Self Homography Learning." pith.science (2026). https://pith.science/paper/KNXFG2AC

@misc{pith2026250608784,
  author       = {Pith},
  title        = {Pith review of: HomographyAD: Deep Anomaly Detection Using Self Homography Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KNXFG2AC}},
  note         = {Machine review of arXiv:2506.08784}
}
read the original abstract

Anomaly detection (AD) is a task that distinguishes normal and abnormal data, which is important for applying automation technologies of the manufacturing facilities. For MVTec dataset that is a representative AD dataset for industrial environment, many recent works have shown remarkable performances. However, the existing anomaly detection works have a limitation of showing good performance for fully-aligned datasets only, unlike real-world industrial environments. To solve this limitation, we propose HomographyAD, a novel deep anomaly detection methodology based on the ImageNet-pretrained network, which is specially designed for actual industrial dataset. Specifically, we first suggest input foreground alignment using the deep homography estimation method. In addition, we fine-tune the model by self homography learning to learn additional shape information from normal samples. Finally, we conduct anomaly detection based on the measure of how far the feature of test sample is from the distribution of the extracted normal features. By applying our proposed method to various existing AD approaches, we show performance enhancement through extensive experiments.

Figures

Figures reproduced from arXiv: 2506.08784 by the authors.

Figure 1
Figure 1. Examples of (Left) public MVTec dataset and (Right) real-world industrial dataset. The images in each column are normal images of the same class. The red outline in the right two columns represents the foreground, the target of the anomaly detection, and the rest represent the background. inspection purposes, resulting in sophisticated yet recurrent patterns as shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of self homography learning process. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visualization of images from the MVtec dataset (1st row) and their trans￾formed images (2nd row). Images of 2nd row are generated such that the images of 1st row are transformed from red outline to green one. 3.1 Input Foreground Alignment For input foreground alignment, we employ the deep homography estimation (DHE) method. Homography estimation aims to find 2D transformation that can be mapped to different images … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Examples of the MVTec dataset. (a) Image samples of original MVTec dataset. Screw, grid, cable, capsule, hazelnut, metalnut, tile, and leather classes in order from top left to bottom right. Left four classes are not aligned (Misaligned Classes of Original MVTec). Righ…
Figure 5
Figure 5. Figure 5: Image samples of texture classes (1st row) and object classes (2nd, 3rd, and [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Visualization of the heatmap for the MVTec dataset. We compared heatmaps from the pre-trained anomaly detection (PAD) approach and PAD w/ homography learning (HL) with the ground truth images. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages

  1. [1]

    Ganomaly: Semi- supervised anomaly detection via adversarial training

    Samet Akcay, Amir Atapour-Abarghouei, and Toby P Breckon. Ganomaly: Semi- supervised anomaly detection via adversarial training. InAsian conference on computer vision, pages 622–637. Springer, 2018

    work page 2018

  2. [2]

    Image based quality inspection in smart manufacturing systems: A literature review.Procedia CIRP, 103:262–267, 2021

    Milica Babic, Mojtaba A Farahani, and Thorsten Wuest. Image based quality inspection in smart manufacturing systems: A literature review.Procedia CIRP, 103:262–267, 2021

    work page 2021

  3. [3]

    Lucas-kanade 20 years on: A unifying frame- work.International journal of computer vision, 56(3):221–255, 2004

    Simon Baker and Iain Matthews. Lucas-kanade 20 years on: A unifying frame- work.International journal of computer vision, 56(3):221–255, 2004

    work page 2004

  4. [4]

    Deep nearest neighbor anomaly detection.arXiv preprint arXiv:2002.10445, 2020

    Liron Bergman, Niv Cohen, and Yedid Hoshen. Deep nearest neighbor anomaly detection.arXiv preprint arXiv:2002.10445, 2020

    arXiv 2002

  5. [5]

    Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 9592–9600, 2019

    work page 2019

  6. [6]

    Clkn: Cascaded lucas- kanade networks for image alignment

    Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas- kanade networks for image alignment. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2213–2221, 2017

    work page 2017

  7. [7]

    Sub-image anomaly detection with deep pyramid correspondences.arXiv preprint arXiv:2005.02357, 2020

    Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences.arXiv preprint arXiv:2005.02357, 2020

    arXiv 2005

  8. [8]

    Padim: a patch distribution modeling framework for anomaly detection and lo- calization

    Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and lo- calization. InInternational Conference on Pattern Recognition, pages 475–489. Springer, 2021

    work page 2021

Show all 35 references
  1. [9]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  2. [10]

    Deep image ho- mography estimation.arXiv preprint arXiv:1606.03798, 2016

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image ho- mography estimation.arXiv preprint arXiv:1606.03798, 2016

    2016 arXiv

  3. [11]

    Homography estimation from image pairs with hierarchical convolutional networks

    Farzan Erlik Nowruzi, Robert Laganiere, and Nathalie Japkowicz. Homography estimation from image pairs with hierarchical convolutional networks. InPro- ceedings of the IEEE international conference on computer vision workshops, pages 913–920, 2017. 15

    2017

  4. [12]

    Imagenet-trained cnns are biased towards tex- ture; increasing shape bias improves accuracy and robustness.arXiv preprint arXiv:1811.12231, 2018

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards tex- ture; increasing shape bias improves accuracy and robustness.arXiv preprint arXiv:1811.12231, 2018

    2018 arXiv

  5. [13]

    Unsupervised represen- tation learning by predicting image rotations.arXiv preprint arXiv:1803.07728, 2018

    Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised represen- tation learning by predicting image rotations.arXiv preprint arXiv:1803.07728, 2018

    2018 arXiv

  6. [14]

    Deep anomaly detection using geometric trans- formations.Advances in neural information processing systems, 31, 2018

    Izhak Golan and Ran El-Yaniv. Deep anomaly detection using geometric trans- formations.Advances in neural information processing systems, 31, 2018

    2018

  7. [15]

    Generative adver- sarial nets.Advances in neural information processing systems, 27, 2014

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde- Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adver- sarial nets.Advances in neural information processing systems, 27, 2014

    2014

  8. [16]

    Deep residual learn- ing for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learn- ing for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  9. [17]

    Using self- supervised learning can improve model robustness and uncertainty.Advances in neural information processing systems, 32, 2019

    Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self- supervised learning can improve model robustness and uncertainty.Advances in neural information processing systems, 32, 2019

    2019

  10. [18]

    Surface defect saliency of magnetic tile.The Visual Computer, 36(1):85–96, 2020

    Yibin Huang, Congying Qiu, and Kui Yuan. Surface defect saliency of magnetic tile.The Visual Computer, 36(1):85–96, 2020

    2020

  11. [19]

    Perceptual loss for robust unsupervised homography estimation

    Daniel Koguciuk, Elahe Arani, and Bahram Zonooz. Perceptual loss for robust unsupervised homography estimation. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 4274–4283, 2021

    2021

  12. [20]

    Distinctive image features from scale-invariant keypoints.Inter- national journal of computer vision, 60(2):91–110, 2004

    David G Lowe. Distinctive image features from scale-invariant keypoints.Inter- national journal of computer vision, 60(2):91–110, 2004

    2004

  13. [21]

    On the generalized distance in statistics

    Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936

    1936

  14. [22]

    Vt-adl: A vision transformer network for image anomaly detection and localization

    Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision transformer network for image anomaly detection and localization. In2021 IEEE 30th International Symposium on In- dustrial Electronics (ISIE), pages 01–06. IEEE, 2021

    2021

  15. [23]

    Unsupervised deep homography: A fast and robust homography estimation model.IEEE Robotics and Automation Letters, 3(3):2346–2353, 2018

    Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model.IEEE Robotics and Automation Letters, 3(3):2346–2353, 2018

    2018

  16. [24]

    GF Page. Multiple view geometry in computer vision, by richard hartley and andrew zisserman, cup, cambridge, uk, 2003, vi+ 560 pp., isbn 0-521-54051- 8.(paperback£ 44.95).Robotica, 23(2):271–271, 2005

    2003

  17. [25]

    Deep learning for anomaly detection: A review.ACM Computing Surveys (CSUR), 54(2):1–38, 2021

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review.ACM Computing Surveys (CSUR), 54(2):1–38, 2021

    2021

  18. [26]

    Modeling the distribution of normal data in pre-trained deep features for anomaly detection

    Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In2020 25th In- 16 ternational Conference on Pattern Recognition (ICPR), pages 6726–6733. IEEE, 2021

    2021

  19. [27]

    Towards total recall in industrial anomaly detection

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

    2022

  20. [28]

    Orb: An ef- ficient alternative to sift or surf

    Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An ef- ficient alternative to sift or surf. In2011 International conference on computer vision, pages 2564–2571. Ieee, 2011

    2011

  21. [29]

    Deep one-class classification

    Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel M¨uller, and Marius Kloft. Deep one-class classification. InInternational conference on machine learning, pages 4393–

  22. [30]

    Unsupervised anomaly detection with generative ad- versarial networks to guide marker discovery

    Thomas Schlegl, Philipp Seeb ¨ock, Sebastian M Waldstein, Ursula Schmidt- Erfurth, and Georg Langs. Unsupervised anomaly detection with generative ad- versarial networks to guide marker discovery. InInternational conference on information processing in medical imaging, pages 1...

    2017

  23. [31]

    Learning and evaluating representations for deep one-class classification.arXiv preprint arXiv:2011.02578, 2020

    Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification.arXiv preprint arXiv:2011.02578, 2020

    2011 arXiv

  24. [32]

    Efficientnet: Rethinking model scaling for convolu- tional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolu- tional neural networks. InInternational conference on machine learning, pages 6105–6114. PMLR, 2019

    2019

  25. [33]

    Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

    2016 arXiv

  26. [34]

    Rethinking planar homography estimation using perspective fields

    Rui Zeng, Simon Denman, Sridha Sridharan, and Clinton Fookes. Rethinking planar homography estimation using perspective fields. InAsian Conference on Computer Vision, pages 571–586. Springer, 2018

    2018

  27. [35]

    Content-aware unsupervised deep homography estimation

    Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. InEuropean Conference on Computer Vision, pages 653–669. Springer, 2020. 17

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.