Beyond Normal References: Discriminative Few-Shot Anomaly Detection

Guansong Pang; Huan Wang; Jun Shen; Jun Yan

arxiv: 2605.23231 · v2 · 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 2026-06-30 16:24 UTC · model grok-4.3

classification 💻 cs.CV

keywords few-shot anomaly detectiondiscriminative anomaly detectionintrinsic deviation learningnormal variation erasureorthogonal deviation directionsgeneralizable abnormalitycomputer visionoutlier detection

0 comments

The pith

IDEAL learns intrinsic deviation patterns from both normal and anomalous few-shot references to detect generalizable anomalies.

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

The paper addresses a discriminative few-shot anomaly detection setting where a small number of both normal and anomalous examples serve as references at inference time. Existing approaches either match only against normal references, missing useful signals from anomalies, or fit both types directly and overfit to the seen anomalies. IDEAL decomposes the process with a Normal Variation Eraser that removes nuisance normal variations and an Intrinsic Deviation Encoder that extracts orthogonal deviation directions from the cleaned representations. At test time the method projects queries onto these directions and scores the remaining deviation from normality. Experiments across eight datasets show improved handling of both seen and unseen anomalies.

Core claim

IDEAL is an intrinsic deviation learning framework that leverages both reference types to learn intrinsic deviation patterns characterizing generalizable abnormality as deviations from normality. It decomposes the learning into a Normal Variation Eraser to suppress nuisance normal variations that may lead to noisy deviations and an Intrinsic Deviation Encoder to decompose the denoised 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 vectors.

What carries the argument

The Intrinsic Deviation Encoder that extracts the most discriminative orthogonal deviation directions from denoised deviation representations after normal variation erasure.

If this is right

The approach outperforms existing state-of-the-art FSAD methods on eight real-world datasets.
It generalizes effectively to unseen anomalies without overfitting to the provided anomalous references.
Scoring relies on deviations that survive projection onto the learned intrinsic directions.
Both normal and anomalous references contribute discriminative information during inference.

Where Pith is reading between the lines

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

The separation of normal variation from anomaly signals could be tested as a modular component in other few-shot detection pipelines.
If the learned directions prove stable across datasets, the method might reduce reliance on collecting many anomaly examples in new domains.
Extending the encoder to handle streaming reference updates could be examined without retraining the full model.

Load-bearing premise

The orthogonal deviation directions extracted from the reference set represent intrinsic generalizable abnormality rather than reference-set-specific artifacts or noise.

What would settle it

Running IDEAL on a new test set containing anomaly types whose deviation directions lie outside the orthogonal subspace learned from the given references and observing whether its advantage over normal-only baselines disappears.

Figures

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

**Figure 1.** Figure 1: Anomaly score maps of (a) an input image using direct similarity matching of (b) anomalous-only references and (c) both anomalous and normal references, compared to (d) our method IDEAL under a 1-shot normal and anomalous reference setting. Two simple reference matching methods produce noisy and spurious anomaly responses, whereas IDEAL yields substantially cleaner activations. Note that, to illustrate … view at source ↗

**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: T-SNE visualization for different 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: (a) Qualitative results of three generalist FSAD methods. (b) Visualization of anomalous 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.

Desk Editor's Note private letter to a colleague

IDEAL adds a clean decomposition for using both normal and anomalous references in few-shot anomaly detection and reports gains across eight datasets, but the key claim about generalizable intrinsic directions rests on an assumption that the experiments need to verify.

read the letter

The main takeaway is that this paper gives a practical way to handle the discriminative FSAD setting by first removing normal variations and then extracting orthogonal deviation directions from the cleaned signals. It then scores queries by how much they deviate along those directions after projection.

What is new is the explicit two-part split into the Normal Variation Eraser and the Intrinsic Deviation Encoder. The eraser targets nuisance factors that could create noisy deviations, while the encoder decomposes the rest into orthogonal vectors meant to capture the most discriminative abnormality patterns. This is a direct response to the limits of pure normality matching or naive fitting of both reference types. The inference procedure follows logically from the decomposition.

The experiments are presented as extensive, covering eight real-world datasets with claims of consistent outperformance and better handling of unseen anomalies. That is the part that would interest people working on industrial or medical vision tasks where a few labeled anomalies are available at test time.

The soft spot is whether the orthogonal directions truly isolate generalizable abnormality rather than patterns specific to the small reference anomaly set. Even after variation erasure, a handful of anomalies can share low-level artifacts or biases that the decomposition picks up. The abstract flags the overfitting risk but does not show a derivation or ablation that proves the orthogonality step removes reference dependence. If the full paper has targeted controls or analysis on this point, it would address the main concern; otherwise the generalization result stays partly assumptive.

I would bring this to a reading group for the framework itself. It is worth sending for peer review because the motivation is clear, the components are well-motivated, and the reported results are substantial enough to merit checking the details and ablations.

Referee Report

2 major / 1 minor

Summary. The paper proposes IDEAL, a framework for discriminative few-shot anomaly detection that uses both normal and anomalous reference examples. It introduces a Normal Variation Eraser to suppress nuisance normal variations and an Intrinsic Deviation Encoder to decompose denoised deviation representations into orthogonal intrinsic deviation vectors. At inference, queries are scored by deviations preserved after projection onto these vectors, with the claim that this enables generalization to both seen and unseen anomalies. Extensive experiments on eight real-world datasets are reported to show consistent outperformance over existing FSAD methods.

Significance. If the central claim holds, the work addresses a practical FSAD setting by incorporating discriminative information from anomalous references without direct overfitting, potentially improving detection performance in data-scarce scenarios. The planned release of code and data would support reproducibility and further validation.

major comments (2)

[Method (Intrinsic Deviation Encoder and inference scoring)] The core assumption that the orthogonal deviation vectors learned by the Intrinsic Deviation Encoder isolate generalizable abnormality (rather than reference-set-specific artifacts or spurious correlations from the few anomalous examples) is load-bearing for the generalization-to-unseen-anomalies claim, yet the manuscript provides no derivation, proof, or targeted ablation demonstrating that orthogonality combined with the Normal Variation Eraser removes reference dependence.
[Experiments] The abstract states that experiments on eight datasets demonstrate outperformance and generalization, but the provided text supplies no details on the specific baselines, evaluation metrics, statistical significance tests, or ablation studies; without these, the empirical support for the central claim cannot be assessed.

minor comments (1)

[Inference] Notation for the projection and scoring steps at inference could be clarified with an explicit equation to avoid ambiguity in how query-to-normal deviations are computed after projection.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and indicate planned revisions.

read point-by-point responses

Referee: [Method (Intrinsic Deviation Encoder and inference scoring)] The core assumption that the orthogonal deviation vectors learned by the Intrinsic Deviation Encoder isolate generalizable abnormality (rather than reference-set-specific artifacts or spurious correlations from the few anomalous examples) is load-bearing for the generalization-to-unseen-anomalies claim, yet the manuscript provides no derivation, proof, or targeted ablation demonstrating that orthogonality combined with the Normal Variation Eraser removes reference dependence.

Authors: We acknowledge that the manuscript does not contain a formal mathematical derivation or proof establishing that orthogonality plus the Normal Variation Eraser fully eliminates reference-set dependence. The framework is motivated by the empirical objective of extracting the most discriminative orthogonal directions in deviation space. In the revision we will add a targeted ablation that systematically varies the anomalous reference sets, measures consistency of the learned vectors across different reference choices, and reports performance on held-out unseen anomalies to empirically support reduced reference dependence. revision: yes
Referee: [Experiments] The abstract states that experiments on eight datasets demonstrate outperformance and generalization, but the provided text supplies no details on the specific baselines, evaluation metrics, statistical significance tests, or ablation studies; without these, the empirical support for the central claim cannot be assessed.

Authors: The full manuscript contains a complete Experiments section (Section 4) that specifies the eight datasets, the full list of baselines, the evaluation metrics (AUROC and AUPRC), multiple-run statistics, and component ablations. If these details were not visible in the version provided for review, we will ensure the section is self-contained and add a concise results summary table to the introduction for immediate visibility. revision: partial

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper proposes IDEAL, a new algorithmic framework with two components (Normal Variation Eraser and Intrinsic Deviation Encoder) that process reference data to produce deviation vectors for scoring. All performance claims rest on empirical results across eight datasets rather than any closed-form derivation or equation that reduces the output to a fitted parameter or self-referential definition. No load-bearing uniqueness theorems, self-citations, or ansatzes are invoked to justify the central method. The approach is self-contained and externally falsifiable via the reported experiments.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

The central claim rests on the introduction of two new algorithmic modules whose effectiveness is asserted via empirical results; no free parameters, mathematical axioms, or invented physical entities are described in the abstract.

invented entities (2)

Normal Variation Eraser no independent evidence
purpose: Suppress nuisance normal variations to highlight anomaly-relevant deviations
New component introduced by the paper
Intrinsic Deviation Encoder no independent evidence
purpose: Decompose denoised deviations into intrinsic orthogonal deviation vectors
New component introduced by the paper

pith-pipeline@v0.9.1-grok · 5762 in / 996 out tokens · 43607 ms · 2026-06-30T16:24:22.806699+00:00 · methodology

Review history (2 revisions) →

Beyond Normal References: Discriminative Few-Shot Anomaly Detection

Core claim

What carries the argument

If this is right

Where Pith is reading between the lines

Load-bearing premise

What would settle it

discussion (0)