ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection¶

Conference: ECCV 2026
Paper: Official paper page · Paper PDF
Code: https://github.com/LGC-AD/ArcAD
Area: Anomaly Detection
Keywords: Industrial anomaly detection, cold start, hyperspherical prototypes, optimal transport, defect-guided calibration

TL;DR¶

ArcAD calibrates the latent space of a reconstruction-based industrial anomaly detector using balanced normal prototypes, prototype-filtered synthetic anomalies, and a few real defects, raising Dinomaly's image-level AUROC from 88.8% to 92.5% on multi-class cold-start Real-IAD while retaining its original reconstruction-based inference score.

Background & Motivation¶

Industrial anomaly detection often learns normal product appearance and identifies departures such as scratches, cracks, or structural defects. Reconstruction-based methods including Dinomaly, RD4AD, and ReContrast compare original and reconstructed features without needing an exhaustive catalog of defect types. However, a new production line may offer too few normal images to represent every texture, material, and local structure. Reconstruction loss then shapes the normal region only indirectly, leaving a poorly organized latent space and an uncertain distinction between unfamiliar normal variation and genuine defects.

A few labeled defects provide useful additional evidence, but training a conventional binary classifier can overfit those particular examples. ArcAD aims to preserve reconstruction-based detection while using scarce anomaly supervision to improve the representation. Its cold-start setting is not a one-support-image protocol: it substantially reduces normal training data and adds a small proportion of genuine anomalies. The central problem is therefore how to organize the observed normal modes and use limited defects to identify directions that should lie outside them.

The paper addresses this geometrically, normalizing patch features onto a unit hypersphere and representing normal appearance with multiple prototypes. Anomaly supervision becomes a training-time constraint rather than a replacement classifier at deployment. Core idea: pull normal features toward balanced prototypes, push real anomalies away from their closest normal prototypes, and align them with filtered synthetic anomalies, changing representation learning without changing reconstruction-based scoring.

Method¶

Overall Architecture¶

Training inputs are limited normal images and a few defective images with localization masks; the output remains the backbone's reconstruction-based anomaly map. Normal images pass through an encoder, bottleneck, and decoder, while ArcAD adds constraints to the bottleneck's patch representations. Its training components are balanced prototype modeling, prototype-restricted synthesis, defect-guided calibration, and auxiliary discrimination. Defect masks select genuinely abnormal patches so that normal background inside a defective image is not treated as anomalous.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Normal images and<br/>masked defect images"] --> B["Encoder and<br/>bottleneck features"]
    B --> C["Balanced prototype modeling"]
    C --> D["Prototype-restricted synthesis"]
    D --> E["Defect-guided calibration"]
    B -->|"Mask-selected real defects"| E
    E --> F["Auxiliary discrimination"]
    F -.->|"Training-time bottleneck update"| B
    B -->|"Original reconstruction path"| G["Decoder and reconstruction error<br/>Inference anomaly map"]

The prototypes, synthesis procedure, and discriminator do not supply the final inference score. The feedback edge denotes optimization, not repeated test-time detection passes. “Plug-and-play” here means integration with reconstruction-based latent representations; the experiments do not establish compatibility with every anomaly detection architecture.

Key Designs¶

1. Balanced prototype modeling: preserve multiple normal modes without assignment collapse

Each bottleneck patch feature undergoes \(\ell_2\) normalization, making comparisons depend on direction rather than feature magnitude. The von Mises–Fisher (vMF) distribution provides the directional clustering interpretation: a prototype is a mean direction, and concentration controls how tightly features gather around it. Multiple prototypes are needed because normal products contain diverse textures and structures, especially when one model handles several product categories. Ordinary clustering can nevertheless allocate most capacity to dense observed regions when training data are scarce.

Sinkhorn-based Prototype Modeling (SPM) initializes prototypes with cosine-distance K-means and casts subsequent soft assignment as balanced optimal transport. The matrix \(Q\) assigns patches to prototypes while favoring high cosine similarity and satisfying:

\[ \sum_{k=1}^{K} Q_{ik}=1,\qquad \sum_{i=1}^{N_{\mathrm{batch}}}Q_{ik}=\frac{N_{\mathrm{batch}}}{K}. \]

Every patch distributes a total weight of one, and every prototype receives the same total assignment mass within the batch. This balances prototype assignments, not the number of images in each product category. Sinkhorn–Knopp iterations rescale an exponentiated similarity matrix to obtain the constrained soft assignments. A cross-entropy objective aligns feature predictions with these assignments, while exponential moving averages update the prototypes in an EM-style alternating procedure. Prototype count supplies multimodal capacity; balanced assignment prevents a few prototypes from monopolizing the observed features. Neither property proves that normal modes absent from the training data have actually been recovered.

2. Prototype-restricted synthesis: reject synthetic defects that still resemble a normal mode

Real defects are too scarce to provide a stable negative region alone, so ArcAD perturbs normal features to create additional negatives. For each normal anchor, it draws five independent Gaussian perturbations and renormalizes every candidate onto the unit hypersphere. Renormalization prevents a candidate from becoming an easy outlier merely through increased magnitude, but does not ensure that its direction has left normal appearance. Labeling all such candidates as anomalies can incorrectly repel valid normal features.

For each candidate, the method computes its maximum cosine similarity to any normal prototype, then retains the candidate with the smallest such maximum. The selected feature must therefore avoid resemblance to all learned normal modes, not just the prototype associated with its starting anchor. Exactly one of the five candidates survives per anchor. This selects the relatively least normal-looking direction among local perturbations; it does not search the entire sphere for a maximally distant point. Nor does it guarantee a fixed boundary distance or select the candidate closest to the normal boundary.

3. Defect-guided calibration: use real defects to identify directions normal features should exclude

Defect-Guided Calibration (DGC) downsamples each defective image's ground-truth mask to patch resolution, extracts only abnormal patches, and normalizes them into the same hyperspherical space. This turns an image-level defect example into local supervision without pushing the image's normal background away from normal prototypes. Consequently, the supervision requirement includes defect masks, not merely a few anomalous-image labels.

Calibration has two complementary effects: reduce each real anomaly's similarity to its most similar normal prototype, and encourage alignment between real anomalies and the synthetic anomaly set. The first separates defects from their most confusable normal mode; the second gives sparse genuine defects a larger, more coherent set of negative examples. The objective is to refine normal representation geometry rather than enumerate every possible semantic defect category. However, drawing heterogeneous defects toward a shared anomalous region is still an assumption, not a universally established property of industrial anomalies.

4. Auxiliary discrimination: strengthen the geometry without replacing the inference rule

A lightweight discriminator, such as a two-layer MLP, labels normal features as zero and both real and synthetic anomalies as one. Binary Focal Loss addresses training imbalance and adds a global separation constraint to the representation. SPM and DGC provide the geometric structure on which this classifier regularizes; the classifier is not expected to discover a reliable boundary from scarce labels alone. Indeed, adding only the discriminator to Dinomaly reduces performance in the ablation study, making its interaction with the other constraints essential.

Loss & Training¶

The bottleneck receives normal reconstruction, SPM, DGC, and classification objectives; each of the three auxiliary loss weights is 0.1. The decoder is updated only through normal-image reconstruction loss, the discriminator only through classification loss, and the architecture diagram marks the encoder as frozen. Real defects are therefore not reconstruction targets that the decoder is encouraged to reproduce as normal. Training uses AdamW for 10,000 iterations per dataset, a learning rate of \(2\times10^{-3}\), momentum parameters \((0.9,0.999)\), and weight decay \(10^{-4}\). The selected prototype count is \(K=500\), reflecting the image/pixel performance tradeoff rather than a shared optimum for every metric.

Equations (2)–(10) in the local text extraction contain damaged mathematical notation, including incomplete operators in the cross-entropy and DGC objectives. Only the assignment constraints explicitly supported by the prose are reproduced here; the verified optimization directions are explained without inventing missing loss expressions. The available cache contains the main paper and references, not the separately cited supplement, so missing per-category results and implementation details are not filled in.

Key Experimental Results¶

Main Results¶

The selected comparison comes from Table 1: a single model is trained across all categories within each dataset, with ArcAD built on Dinomaly. The cold-start protocol retains 30% of the original normal training set and adds anomalies sampled from the original test set. Anomalies normally form 10% of the resulting training set, reduced to 5% for VisA and MANTA. Removed normal training images enter the new test set, while anomalies used for training leave it; all methods use that same reconstructed test split. Unsupervised Dinomaly receives only normal training images, whereas ArcAD additionally receives anomalies and their masks, so the supervision budgets are not equal. Scores are class averages in percent, and all three metrics are higher-is-better. I-AUROC measures image-level ranking, P-AUROC measures pixel-level ranking, and P-F1-max is the best pixel F1 across thresholds rather than F1 at a fixed deployment threshold.

Dataset	Dinomaly: I-AUROC / P-AUROC / P-F1-max	ArcAD: I-AUROC / P-AUROC / P-F1-max	I-AUROC gain, percentage points
MVTec-AD	99.6 / 98.3 / 68.4	99.7 / 99.2 / 68.9	+0.1
VisA	98.4 / 98.9 / 53.9	98.9 / 99.0 / 54.9	+0.5
Real-IAD	88.8 / 98.8 / 46.7	92.5 / 99.0 / 49.8	+3.7
MANTA tiny	90.3 / 93.8 / 39.8	93.3 / 95.5 / 48.5	+3.0

Source: Table 1; split construction is described in Section 4.1, which specifies the tiny version of MANTA. The starred Dinomaly row in the original table is a standard-setting reference, not the same-setting baseline above. For example, its MANTA P-F1-max is 53.1 versus ArcAD's 48.5, so the paper's broad performance language should not be expanded into superiority over every reference number in the table.

Ablation Study¶

All rows below use multi-class cold-start Real-IAD; scores are percentages and higher is better. The table selects component ablations from Table 4 and one especially informative synthesis ablation from Table 6, with sources distinguished explicitly.

Configuration	I-AUROC	P-AUROC	P-F1-max	Source and interpretation
Dinomaly baseline	88.8	98.8	46.7	Table 4; reconstruction only
Discriminator only added	87.8	98.5	45.5	Table 4; no SPM or DGC
Without SPM	91.7	98.9	49.3	Table 4; omit prototype-modeling loss
Without DGC	90.2	98.8	48.1	Table 4; omit calibration loss
Without candidate filtering	90.2	98.5	43.4	Table 6; use all synthetic candidates
Full ArcAD	92.5	99.0	49.8	Tables 4 and 6

Key Findings¶

Removing DGC costs 2.3 I-AUROC points and 1.7 P-F1-max points; removing SPM costs 0.8 and 0.5 points, respectively. Defect-guided geometry supplies a substantial part of the measured benefit.
Candidate filtering particularly affects localization: P-F1-max rises from 43.4 without filtering to 49.8 with the complete method, a 6.4-point difference. More synthetic negatives do not automatically mean better supervision.
Additional labels alone are insufficient: adding only the discriminator lowers I-AUROC from 88.8 to 87.8. The gain depends on how supervision shapes the representation.
Table 3 in Section 4.2 reports I-AUROC of 89.7 for Dinomaly plus ArcAD with a 3% anomaly ratio on Real-IAD, above the 88.8 cold-start baseline but below 92.5 with 10% anomalies.

Highlights & Insights¶

Training supervision and inference scoring are deliberately separated. Defects can explicitly guide representation learning while deployment keeps the familiar reconstruction discrepancy.
Normal prototypes also act as synthetic-negative quality checks. They organize normal appearance and reject candidates close to any learned normal mode, giving both operations a common geometric reference.
Multiple prototypes and balanced assignment address distinct failure modes. Capacity handles multimodal appearance, while transport constraints prevent scarce observations from concentrating assignments in a few prototypes.

Limitations & Future Work¶

The authors acknowledge additional training computation from prototype optimization and calibration, and propose lighter implementations. Unchanged inference scoring does not mean unchanged training cost.
Reader assessment: retaining 30% of normal training images is not equivalent to a fixed one-shot or few-shot support set. Results should not be generalized to every low-data protocol.
Reader assessment: genuine defects require pixel masks. A small anomaly count does not remove the associated localization-annotation cost.
Reader assessment: the main tables do not report variation across random splits or confidence intervals. The 0.1-point MVTec-AD image-level gain should not be treated as an established statistically stable improvement.
Reader assessment: equal prototype mass need not match the long-tailed frequencies of real normal modes. The experiments establish empirical usefulness here, not universal suitability under severe mode imbalance.

Compared with Dinomaly, RD4AD, and ReContrast: these provide the reconstruction backbone and inference score; ArcAD adds latent-space training constraints rather than a new test-time scoring mechanism.
Compared with SimpleNet and DRAEM: these represent feature-space and image-space anomaly synthesis, respectively. ArcAD instead renormalizes perturbations onto a hypersphere, filters them against all normal prototypes, and combines them with localized genuine defects.
Compared with supervised methods such as DRA: ArcAD emphasizes using known defects to rectify normal boundaries rather than relying only on representations discriminating observed anomalies. Backbone and optimization differences nevertheless prevent attributing the entire performance gap to that conceptual distinction.

Rating¶

Novelty: 3.5/5. The underlying tools are established, but their integration and supervision interface are tailored to cold-start reconstruction models.
Experimental Thoroughness: 4/5. Four datasets, single-/multi-class settings, multiple reconstruction backbones, and useful ablations provide broad coverage; statistical variability is not reported.
Writing Quality: 3.5/5. The main mechanism is clear, although broad performance claims require care around standard-setting reference rows; extraction damage is not counted as a defect in the original writing.
Value: 4/5. A practical training-time enhancement for reconstruction-based inspection systems that can obtain a few masked defects without changing the final scoring path.