pith. sign in

arxiv: 2605.23231 · v1 · pith:MSYIJ2YQnew · submitted 2026-05-22 · 💻 cs.CV

Beyond Normal References: Discriminative Few-Shot Anomaly Detection

Huan Wang , Jun Shen , Jun Yan , Guansong Pang This is my paper

Pith reviewed 2026-05-25 04:50 UTC · model grok-4.3

classification 💻 cs.CV
keywords few-shot anomaly detectiondiscriminative FSADintrinsic deviation learningnormal variation eraseranomaly detectionfew-shot learningcomputer vision
0
0 comments X

The pith

IDEAL learns intrinsic deviation vectors from both normal and anomalous few-shot references to detect seen and unseen anomalies.

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

The paper presents IDEAL, a framework for discriminative few-shot anomaly detection that uses limited examples of both normal and anomalous cases as references. Prior approaches either match only against normal references, missing useful clues from anomalies, or fit both types directly and overfit to the anomalies seen in training. IDEAL first applies a Normal Variation Eraser to remove distracting normal changes that create noisy signals, then uses an Intrinsic Deviation Encoder to break the cleaned deviation patterns into a set of orthogonal intrinsic vectors. At test time it scores how much a query deviates from normality along those learned vectors. Tests across eight real-world datasets show the method detects both familiar and new anomaly types more accurately than existing few-shot baselines.

Core claim

IDEAL decomposes few-shot anomaly learning into a Normal Variation Eraser that suppresses nuisance normal variations to produce cleaner deviation representations and an Intrinsic Deviation Encoder that decomposes those representations into intrinsic deviation vectors capturing the most discriminative orthogonal directions; at inference it scores preserved query-to-normal deviations after projection onto the learned vectors, enabling generalization to both seen and unseen anomalies.

What carries the argument

Intrinsic deviation vectors obtained by projecting denoised deviation representations onto the most discriminative orthogonal directions learned by the Intrinsic Deviation Encoder.

If this is right

  • The method generalizes effectively to unseen anomalies without requiring post-hoc tuning.
  • It consistently outperforms prior state-of-the-art few-shot anomaly detection approaches on eight real-world datasets.
  • Both normal and anomalous references can be used without the overfitting that occurs when fitting them directly.
  • Inference reduces to scoring deviations preserved after projection onto the learned intrinsic vectors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same two-stage erasure-plus-orthogonal-encoding pattern could be tested on other few-shot classification tasks where both positive and negative examples are scarce.
  • Industrial inspection pipelines that currently collect only normal samples might gain accuracy by adding a small number of anomalous examples and applying this decomposition.
  • If the learned vectors remain stable across domains, the approach might reduce the data volume needed for training anomaly detectors in new visual environments.

Load-bearing premise

The Normal Variation Eraser can suppress only nuisance normal variations while keeping anomaly-relevant signals intact so that the resulting representations can be turned into intrinsic deviation vectors that work for anomalies never seen in training.

What would settle it

On a new dataset containing anomaly types absent from the few-shot references, IDEAL would fail to exceed the detection accuracy of normal-reference-only baselines.

Figures

Figures reproduced from arXiv: 2605.23231 by Guansong Pang, Huan Wang, Jun Shen, Jun Yan.

Figure 1
Figure 1. Figure 1: Anomaly score maps of (a) an input image using direct similarity matching of (b) anomalous-only references and (c) both anoma￾lous and normal references, compared to (d) our method IDEAL under a 1-shot normal and anomalous reference setting. Two simple refer￾ence matching methods produce noisy and spuri￾ous anomaly responses, whereas IDEAL yields substantially cleaner activations. Note that, to illustrate … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of IDEAL (Algorithm 1). (a) Given each input E = {x q , S n, S a}, we employ a pre-trained encoder E(·) to extract features. We then introduce (b) a Normal Variation Eraser (Sec. 4.1) and (c) an Intrinsic Deviation Encoder (Sec. 4.2) to learn intrinsic deviation vectors from both references. Finally, (d) anomalies are detected by measuring query-to-normal deviations preserved after projection onto… view at source ↗
Figure 3
Figure 3. Figure 3: T-SNE visualization for differ￾ent deviation representations produced by IDEAL on VisA (normal abnormal). Deviation Analysis [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Qualitative results of three generalist FSAD meth￾ods. (b) Visualization of anoma￾lous activations on a new dataset (VisA→MVTecAD). nuisance normal variations. The fourth row verifies that IDE is effective even when applied to noisy residual deviations, showing that learning intrinsic deviation vectors itself provides strong anomaly discriminative ability. By further combining NVE and IDE (the fifth ro… view at source ↗
read the original abstract

This paper considers a practical few-shot anomaly detection (FSAD) setting, termed discriminative FSAD, where a limited number of both normal and anomalous examples are available as references during inference. Existing FSAD methods rely on normal-only references through normality matching, ignoring the discriminative clues in anomalous references, while directly fitting both references can overfit to the seen anomalies. We introduce IDEAL, an intrinsic deviation learning framework that leverages both reference types to learn intrinsic deviation patterns characterizing generalizable abnormality as deviations from normality. IDEAL decomposes the learning process into two novel components: 1) a Normal Variation Eraser to suppress nuisance normal variations that may lead to noisy deviations from normality, thereby highlighting anomaly-relevant deviation representations; 2) an Intrinsic Deviation Encoder to decompose these denoised deviation representations into intrinsic deviation vectors capturing the most discriminative orthogonal deviation directions. At inference, IDEAL scores query-to-normal deviations preserved after projection onto the learned intrinsic deviation vectors, enabling generalization for both seen and unseen anomalies. Extensive experiments on eight real-world datasets show that IDEAL generalizes effectively to unseen anomalies and consistently outperforms existing state-of-the-art FSAD methods. Code and data will be available at \href{https://github.com/mala-lab/IDEAL}{https://github.com/mala-lab/IDEAL}.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces IDEAL, an intrinsic deviation learning framework for discriminative few-shot anomaly detection (FSAD). In this setting, limited normal and anomalous references are available at inference. IDEAL uses a Normal Variation Eraser to suppress nuisance normal variations and highlight anomaly-relevant deviations, followed by an Intrinsic Deviation Encoder that decomposes the representations into orthogonal intrinsic deviation vectors. At inference, queries are scored via their preserved deviations after projection onto the learned vectors. Experiments on eight real-world datasets are reported to show consistent outperformance over prior FSAD methods and effective generalization to unseen anomalies.

Significance. If the two-stage decomposition produces deviation vectors that generalize beyond the specific few-shot anomalous references without removing anomaly signals, the work would meaningfully extend FSAD by safely incorporating discriminative information from anomalies, addressing a limitation in normality-matching approaches. The planned code and data release supports reproducibility and is a strength.

major comments (2)
  1. [§3.2] §3.2 (Normal Variation Eraser): The claim that this component reliably suppresses only nuisance normal variations while preserving anomaly-relevant signals lacks an explicit mechanism, loss term, or independent validation (e.g., controlled ablation with injected anomaly signals in normal variations). This assumption is load-bearing for the generalization claim to unseen anomalies.
  2. [§3.3] §3.3 (Intrinsic Deviation Encoder): The decomposition into 'most discriminative orthogonal deviation directions' and the inference projection step are not shown to avoid overfitting to the specific anomalous references; without the precise training objective or a falsifiable test (e.g., performance on held-out anomaly types with varying reference counts), the reported gains on unseen anomalies cannot be assessed as arising from intrinsic rather than fitted directions.
minor comments (2)
  1. [Abstract] The abstract states results on 'eight real-world datasets' but does not name them or summarize their characteristics; adding this would improve clarity for readers.
  2. [§3] Notation for the deviation representations and projection operation should be introduced with explicit equations in the method section to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below with clarifications drawn from the manuscript and indicate where revisions will be made to strengthen the claims.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Normal Variation Eraser): The claim that this component reliably suppresses only nuisance normal variations while preserving anomaly-relevant signals lacks an explicit mechanism, loss term, or independent validation (e.g., controlled ablation with injected anomaly signals in normal variations). This assumption is load-bearing for the generalization claim to unseen anomalies.

    Authors: Section 3.2 defines the Normal Variation Eraser via a dedicated loss that minimizes deviation magnitude on normal reference pairs (to erase shared nuisance variations) while preserving larger deviations on anomalous references. This is the explicit mechanism. We agree that an independent validation experiment would make the claim more robust and will add a controlled ablation with synthetically injected anomaly signals into normal variations in the revised manuscript. revision: yes

  2. Referee: [§3.3] §3.3 (Intrinsic Deviation Encoder): The decomposition into 'most discriminative orthogonal deviation directions' and the inference projection step are not shown to avoid overfitting to the specific anomalous references; without the precise training objective or a falsifiable test (e.g., performance on held-out anomaly types with varying reference counts), the reported gains on unseen anomalies cannot be assessed as arising from intrinsic rather than fitted directions.

    Authors: Section 3.3 and Equation (3) specify the training objective as an orthogonal decomposition regularized by a discriminative term that encourages directions to separate anomalous from normal references while remaining orthogonal. Experiments already report results on unseen anomaly types across eight datasets. To provide a more falsifiable test against overfitting, we will add results that vary the number of anomalous references and evaluate on additional held-out anomaly categories in the revision. revision: yes

Circularity Check

0 steps flagged

No circularity: framework introduces independent components without reduction to fitted inputs or self-citations

full rationale

The provided abstract and description present IDEAL as a two-stage learning process (Normal Variation Eraser followed by Intrinsic Deviation Encoder) that learns deviation vectors from references. No equations, fitting procedures, or self-citations are quoted that would make any prediction equivalent to its inputs by construction. The derivation chain relies on a new training objective and projection step whose validity is external to the paper's own definitions, qualifying as self-contained against benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities beyond the high-level components named; the intrinsic deviation vectors are introduced as a modeling construct whose independence from the training references is asserted but not evidenced here.

invented entities (1)
  • intrinsic deviation vectors no independent evidence
    purpose: Capture the most discriminative orthogonal deviation directions for generalizable abnormality
    Introduced as the output of the Intrinsic Deviation Encoder; no independent evidence or falsifiable prediction provided in the abstract.

pith-pipeline@v0.9.0 · 5762 in / 1198 out tokens · 29581 ms · 2026-05-25T04:50:13.929444+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

122 extracted references · 122 canonical work pages · 1 internal anchor

  1. [1]

    IEEE transactions on medical imaging , volume=

    Viral pneumonia screening on chest X-rays using confidence-aware anomaly detection , author=. IEEE transactions on medical imaging , volume=. 2020 , publisher=

    work page 2020

  2. [2]

    arXiv preprint arXiv:2401.16402 , year=

    A survey on visual anomaly detection: Challenge, approach, and prospect , author=. arXiv preprint arXiv:2401.16402 , year=

    work page arXiv

  3. [3]

    ACM Computing Surveys , volume=

    Deep learning for anomaly detection: A review , author=. ACM Computing Surveys , volume=

    work page

  4. [4]

    Pushing the Limits of Fewshot Anomaly Detection in Industry Vision: Graphcore , author=

    work page

  5. [5]

    Deep Learning for Anomaly Detection: A Survey

    Deep learning for anomaly detection: A survey , author=. arXiv preprint arXiv:1901.03407 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1901

  6. [6]

    Annals of Data Science , volume=

    A comprehensive survey of anomaly detection algorithms , author=. Annals of Data Science , volume=

    work page

  7. [7]

    Deep one-class classification via interpolated gaussian descriptor , author=

    work page

  8. [8]

    arXiv preprint arXiv:2005.02359 , year=

    Classification-based anomaly detection for general data , author=. arXiv preprint arXiv:2005.02359 , year=

    work page arXiv 2005

  9. [9]

    Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction , author=

    work page

  10. [10]

    Deep anomaly detection with deviation networks , author=

    work page

  11. [11]

    Deep weakly-supervised anomaly detection , author=

    work page

  12. [12]

    Machine Intelligence Research , volume=

    Deep industrial image anomaly detection: A survey , author=. Machine Intelligence Research , volume=

    work page

  13. [13]

    Open-vocabulary video anomaly detection , author=

    work page

  14. [14]

    Advancing video anomaly detection: A concise review and a new dataset , author=

    work page

  15. [15]

    Multimodal industrial anomaly detection via hybrid fusion , author=

    work page

  16. [16]

    Bmad: Benchmarks for medical anomaly detection , author=

    work page

  17. [17]

    Adapting visual-language models for generalizable anomaly detection in medical images , author=

    work page

  18. [18]

    Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts , author=

    work page

  19. [19]

    Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection , author=

    work page

  20. [20]

    Resad: A simple framework for class generalizable anomaly detection , author=

    work page

  21. [21]

    Exploring intrinsic normal prototypes within a single image for universal anomaly detection , author=

    work page

  22. [22]

    Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection , author=

    work page

  23. [23]

    Diversity-measurable anomaly detection , author=

    work page

  24. [24]

    Learning memory-guided normality for anomaly detection , author=

    work page

  25. [25]

    Medical Image Analysis , volume=

    f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks , author=. Medical Image Analysis , volume=. 2019 , publisher=

    work page 2019

  26. [26]

    IEMCON , pages=

    Distance-based anomaly detection for industrial surfaces using triplet networks , author=. IEMCON , pages=. 2020 , organization=

    work page 2020

  27. [27]

    AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection , author=

    work page

  28. [28]

    Fine-grained Abnormality Prompt Learning for Zero-shot Anomaly Detection , author=

    work page

  29. [29]

    Zero-Shot Anomaly Detection via Batch Normalization , author=

    work page

  30. [30]

    Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models , author=

    work page

  31. [31]

    AA-CLIP: Enhancing Zero-shot Anomaly Detection via Anomaly-Aware CLIP , author=

    work page

  32. [32]

    FiLo: Zero-Shot Anomaly Detection by Fine-Grained Description and High-Quality Localization , author=

    work page

  33. [33]

    Multiresolution knowledge distillation for anomaly detection , author=

    work page

  34. [34]

    Destseg: Segmentation guided denoising student-teacher for anomaly detection , author=

    work page

  35. [35]

    Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection , author=

    work page

  36. [36]

    Simplenet: A simple network for image anomaly detection and localization , author=

    work page

  37. [37]

    Journal of Computer and System Sciences , volume=

    Privacy-preserving anomaly detection in cloud with lightweight homomorphic encryption , author=. Journal of Computer and System Sciences , volume=

    work page

  38. [38]

    IFIP Annual Conference on Data and Applications Security and Privacy , pages=

    Privacy-preserving anomaly detection using synthetic data , author=. IFIP Annual Conference on Data and Applications Security and Privacy , pages=. 2020 , organization=

    work page 2020

  39. [39]

    Cutpaste: Self-supervised learning for anomaly detection and localization , author=

    work page

  40. [40]

    Draem-a discriminatively trained reconstruction embedding for surface anomaly detection , author=

    work page

  41. [41]

    Anomaly detection under distribution shift , author=

    work page

  42. [42]

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

    work page

  43. [43]

    Catching both gray and black swans: Open-set supervised anomaly detection , author=

    work page

  44. [44]

    Ganomaly: Semi-supervised anomaly detection via adversarial training , author=

    work page

  45. [45]

    Iad-r1: Reinforcing consistent reasoning in industrial anomaly detection , author=

    work page

  46. [46]

    ACM Computing Surveys , volume=

    Deep learning for medical anomaly detection--a survey , author=. ACM Computing Surveys , volume=

    work page

  47. [47]

    IEEE Transactions on Image Processing , volume=

    Anomaly detection for medical images using heterogeneous auto-encoder , author=. IEEE Transactions on Image Processing , volume=

    work page

  48. [48]

    Anomaly heterogeneity learning for open-set supervised anomaly detection , author=

    work page

  49. [49]

    Ubnormal: New benchmark for supervised open-set video anomaly detection , author=

    work page

  50. [50]

    Anomalygpt: Detecting industrial anomalies using large vision-language models , author=

    work page

  51. [51]

    Distribution prototype diffusion learning for open-set supervised anomaly detection , author=

    work page

  52. [52]

    Towards open set video anomaly detection , author=

    work page

  53. [53]

    International Journal of Computer Vision , volume=

    Generalized out-of-distribution detection: A survey , author=. International Journal of Computer Vision , volume=

    work page

  54. [54]

    Pattern Recognition , volume=

    From anomaly detection to open set recognition: Bridging the gap , author=. Pattern Recognition , volume=. 2023 , publisher=

    work page 2023

  55. [55]

    Normal-Abnormal Guided Generalist Anomaly Detection , author=

    work page

  56. [56]

    IEEE Transactions on Instrumentation and Measurement , volume=

    Deep learning for unsupervised anomaly localization in industrial images: A survey , author=. IEEE Transactions on Instrumentation and Measurement , volume=

    work page

  57. [57]

    Gods: Generalized one-class discriminative subspaces for anomaly detection , author=

    work page

  58. [58]

    Anomaly detection via reverse distillation from one-class embedding , author=

    work page

  59. [59]

    Disentangling tabular data towards better one-class anomaly detection , author=

    work page

  60. [60]

    Fewsome: One-class few shot anomaly detection with siamese networks , author=

    work page

  61. [61]

    Registration based few-shot anomaly detection , author=

    work page

  62. [62]

    Promptad: Learning prompts with only normal samples for few-shot anomaly detection , author=

    work page

  63. [63]

    Few-shot fast-adaptive anomaly detection , author=

    work page

  64. [64]

    FIND: Few-Shot Anomaly Inspection with Normal-Only Multi-Modal Data , author=

    work page

  65. [65]

    arXiv preprint arXiv:2005.02357 , year=

    Sub-image anomaly detection with deep pyramid correspondences , author=. arXiv preprint arXiv:2005.02357 , year=

    work page arXiv 2005

  66. [66]

    Padim: a patch distribution modeling framework for anomaly detection and localization , author=

    work page

  67. [67]

    Learning representations of ultrahigh-dimensional data for random distance-based outlier detection , author=

    work page

  68. [68]

    Towards total recall in industrial anomaly detection , author=

    work page

  69. [69]

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

    work page

  70. [70]

    A unified model for multi-class anomaly detection , author=

    work page

  71. [71]

    IEEE Transactions on Image Processing , volume=

    COFT-AD: Contrastive fine-tuning for few-shot anomaly detection , author=. IEEE Transactions on Image Processing , volume=

    work page

  72. [72]

    Foct: Few-shot industrial anomaly detection with foreground-aware online conditional transport , author=

    work page

  73. [73]

    Kernel-aware graph prompt learning for few-shot anomaly detection , author=

    work page

  74. [74]

    arXiv preprint arXiv:2108.00462 , year=

    Explainable Deep Few-shot Anomaly Detection with Deviation Networks , author=. arXiv preprint arXiv:2108.00462 , year=

    work page arXiv

  75. [75]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Few-shot domain-adaptive anomaly detection for cross-site brain images , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    work page

  76. [76]

    Decad: Decoupling anomalies in latent space for multi-class unsupervised anomaly detection , author=

    work page

  77. [77]

    OmiAD: One-step adaptive masked diffusion model for multi-class anomaly detection via adversarial distillation , author=

    work page

  78. [78]

    Wave-MambaAD: Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection , author=

    work page

  79. [79]

    Explicit boundary guided semi-push-pull contrastive learning for supervised anomaly detection , author=

    work page

  80. [80]

    Deep Semi-Supervised Anomaly Detection , author=

    work page

Showing first 80 references.