acceptodds
Under review as a conference paper at ICLR 2027

PARC: Joint Modeling of Patch Appearance and Regional Composition for Training-Free Structural and Logical Anomaly Detection

Abstract

In real-world industrial anomaly detection, production lines may contain products exhibiting both structural and logical anomalies. While conventional anomaly detection methods perform well in detecting structural anomalies, they often struggle with logical anomalies. Recent methods alleviate this limitation but often require category-specific training or auxiliary modules, which can limit generalizability and introduce substantial computational overhead. To overcome these limitations, we propose , a training-free framework that jointly models atch ppearance and egional omposition using shared normal references. For appearance assessment, PARC evaluates local deviations by matching patch features against normal references. For composition assessment, it groups normal features into recurring patterns, summarizes each image by the proportions of these patterns within each spatial region, and scores how unusual these regional proportions are relative to normal images. Because structural and logical anomalies may occur independently, PARC calibrates the two scores using normal score distributions and fuses them for effective detection. Without requiring auxiliary modules or category-specific training, PARC achieves a combined image-level AUROC of 91.12% on MVTec-LOCO-AD using frozen features and raises the mean AUROC using five different backbones from 72.69% to 84.74%.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
anomaly detection, patch matching, logical anomalies, score calibration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_uNTgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PARC, a training-free method for industrial anomaly detection. It combines local appearance scores with regional composition scores to detect both structural and logical anomalies. The method uses frozen visual features, K-means clustering, statistical modeling, and score calibration. The experiments show promising results on several industrial anomaly detection benchmarks.

Critical strengths
  1. Strong Performance. The proposed method achieves competitive performance in both structural and logical anomaly detection, outperforming several existing approaches.

  2. High Computational Efficiency. PARC achieves strong detection performance with significantly lower computational cost than methods such as LogSAD, without requiring per-category neural network training.

  3. Simple and Intuitive Framework. The idea of combining local appearance and regional composition modeling is intuitive and well-motivated. The overall framework is easy to follow.

  4. Comprehensive Experiments. The experimental evaluation is relatively thorough, including comparisons with existing methods, evaluations across different feature extractors and datasets, and detailed ablation studies that support the effectiveness of the proposed components.

Critical weaknesses
  1. Limited Methodological Novelty. The proposed method mainly combines existing techniques for feature distribution modeling and anomaly score fusion. SFAD [1] already models the distribution of local elements to detect unusual combinations of normal features, while LogSAD explores score calibration and fusion for structural and logical anomaly detection. Although PARC introduces regional pattern statistics, its main components are largely based on existing techniques, raising concerns about its methodological novelty.

  2. Potential Limitations of Simplified Composition Modeling. Unlike LogSAD, which explicitly models object-level composition rules, PARC represents logical anomalies through regional distributions of local visual patterns. Although this simplified representation offers significant computational advantages, it does not explicitly model object identities or their relationships. Consequently, valid variations in object arrangements may be incorrectly detected as anomalies, while logical violations that preserve similar regional pattern distributions may remain undetected. Although additional experiments are conducted on MVTec AD, VisA, and CAD-SD, the evaluation remains insufficient to demonstrate the effectiveness of this representation on more complex logical anomalies.

  3. Limited Evidence for Spatial Composition Modeling. Although PARC emphasizes regional composition modeling, the ablation results suggest that spatial information contributes relatively modest improvements. Removing regional proportions or permuting patch locations reduces Combined AUROC by only 1.07 and 0.76 points, respectively, whereas replacing soft assignment with hard assignment causes a much larger drop of 4.22 points. These results raise the question of whether the performance gains primarily come from visual pattern statistics rather than spatial composition modeling. Further analysis would help clarify the actual contribution of spatial relationships to logical anomaly detection.

[1] Niv Cohen, et al. Set Features for Anomaly Detection, TMLR, 2024.

Questions for authors
  1. Methodological Contribution. How does PARC differ from existing methods such as ComAD, CSAD, SFAD [1], and LogSAD? In particular, what are the key advantages of using implicit visual prototypes and regional statistics instead of explicit component representations or other composition modeling approaches?

  2. Robustness to Unseen Normal Arrangements. How well can PARC distinguish unseen but valid object arrangements from genuine logical anomalies? For example, evaluating on unseen normal arrangements alongside abnormal configurations and reporting the corresponding false positive rates would help clarify the robustness of the proposed composition modeling.

  3. Contribution of Spatial Information. In Table 5, removing regional proportions or permuting patch locations reduces Combined AUROC by only 1.07 and 0.76 points, respectively, while replacing soft assignment with hard assignment causes a larger drop of 4.22 points. Do these results suggest that PARC primarily benefits from visual pattern statistics rather than spatial composition modeling?

  4. Spatial Matching. How does PARC's global nearest-neighbor matching with position-specific normalization compare with conventional feature matching methods such as PatchCore and spatially aware matching methods such as HETMM [2], particularly in handling normal variations in object positions and orientations?

  5. Score Fusion. In Table 4, the final fusion improves Combined AUROC but reduces Logical AUROC from 90.88 to 89.42 compared with the composition score alone. What causes this degradation, and how does it affect the effectiveness of PARC for logical anomaly detection?

Overall, the paper demonstrates good experimental results, and I appreciate the efficiency of the proposed approach. However, my main concern is the capability and limitations of the simplified composition modeling, particularly whether it can reliably address the diverse and complex logical anomalies encountered in real-world industrial settings. A convincing clarification or additional evidence addressing this concern would positively influence my overall assessment and potentially lead me to increase my rating.

[2] Zixuan Chen, et al. Hard-Normal Example-Aware Template Mutual Matching for Industrial Anomaly Detection, IJCV, 2025.

7 comments
AuthorsOpenReview ↗

Author Responses to Reviewer uNTg (Part 1)

We would like to sincerely thank you for your efforts and valuable comments to improve our work!

Below we address your concerns.


Q1: Method contribution and implicit composition modeling.

How does PARC differ from existing methods such as ComAD, CSAD, SFAD [1], and LogSAD? In particular, what are the key advantages of using implicit visual prototypes and regional statistics instead of explicit component representations or other composition modeling approaches?

A1(part 1):

Thank you for asking us to clarify the contribution.

PARC adds a composition check to local appearance detection using the same frozen features, without component masks, neural-network training for each product category, or explicit product rules. Position-normalized patch matching checks local defects. Continuous responses to normal prototypes—recurring visual feature patterns—describe regional composition. The two scores are calibrated using normal images and combined.

Differences from prior work.

  1. SFAD describes feature sets using random-projection histograms and then applies nearest-neighbor scoring. Discrete bins ignore within-bin variation, and scoring requires retrieving normal descriptors. PARC jointly describes regions using continuous responses to normal prototypes, avoiding hard binning and nearest-neighbor retrieval of composition descriptors.

  2. ComAD first extracts component regions and then measures their areas, colors, and counts. Errors in component boundaries or instance separation may affect subsequent decisions. PARC directly aggregates visual pattern responses without forming complete component regions, reducing dependence on segmentation quality and the cost of region postprocessing.

  3. CSAD first generates component pseudo-labels and then trains a category-specific segmentation network, requiring additional data preparation and training. PARC builds normal composition statistics directly from frozen features, avoiding component pseudo-labels and category-specific network training and reducing adaptation costs for new products.

  4. LogSAD relies on object segmentation, object matching, and product rules, adding inference cost and rule preparation. PARC computes pattern proportions, regional distributions, and their relationships from normal images, avoiding these steps and reducing inference and deployment costs.

MethodExisting designPARC differencePARC advantage
SFADRandom-projection histograms and nearest-neighbor descriptor scoringContinuous prototype responses, joint regional statistics and Mahalanobis scoringNo hard binning or nearest-neighbor retrieval of composition descriptors
ComADComponent regions with area, color and count statisticsRegional visual-pattern responses without explicit component regionsLess dependence on component boundaries and instance separation, reduced postprocessing cost
CSADComponent pseudo-labels and category-specific segmentation trainingNormal statistics fitted directly to frozen featuresNo pseudo-label preparation or category-specific network training, lower adaptation cost
LogSADObject segmentation, object matching and product-specific rulesPattern proportions, regional distributions and correlations fitted from normal imagesNo object segmentation, object matching or product-rule preparation, lower inference and deployment costs

This combination of continuous composition representation, independent local inspection, and shared features changes both the representation and the detection pipeline.

AuthorsOpenReview ↗

Author Responses to Reviewer uNTg (Part 2)

A1(part 2):

Advantages of implicit visual prototypes and regional statistics.

Efficiency is the primary practical advantage. Explicit methods generate component masks before measuring or matching objects. Using a large segmenter such as the Segment Anything Model (SAM) can add substantial latency. PARC instead computes prototype similarities, regional summaries, and statistical distances from existing features. To measure the extra computation, we time composition scoring and fusion. They add 1.83 ms/image. The result shows that composition can be added to appearance detection at low computational cost.

To compare full-system accuracy and runtime, Table 1 reports the default complete systems. ComAD and LogSAD use explicit components. None of these three methods requires target-category network training.

MethodCombined AUROC (%)Latency (ms/image)
ComAD82.01168.39
LogSAD90.242138.66
PARC91.12114.35

The results show that PARC has comparable or higher detection performance at lower system cost. LogSAD takes 18.7 times as long per image.

Implicit representation reduces preparation. Prototypes need not correspond to whole semantic parts: several patterns can describe one component, without specifying its name, precise boundary, or instance assignment. This avoids preparing component definitions, object matching, and product-specific rules. To compare detection without explicit rules under shared features and normal data, we use RADIO, DINOv2, and CLIP, with the same backbone weights and normal training images for both methods within each comparison. Neither method uses explicit rules. PARC outperforms LogSAD with all three backbones by 5.31–9.93 percentage points in Logical AUROC. The results show that PARC provides useful composition evidence without explicit rules.

We have clarified PARC's contributions and advantages in the Introduction (Page 2–3 of the revised manuscript). The representation and efficiency details are provided in Sections 3.3 (Page 5 of the revised manuscript) and 4.1 (Page 8 of the revised manuscript).

[1] Niv Cohen, Issar Tzachor, and Yedid Hoshen. Set Features for Anomaly Detection. TMLR, 2024.

AuthorsOpenReview ↗

Author Responses to Reviewer uNTg (Part 3)

Q2: Unseen valid arrangements and normal-reference coverage.

How well can PARC distinguish unseen but valid object arrangements from genuine logical anomalies? For example, evaluating on unseen normal arrangements alongside abnormal configurations and reporting the corresponding false positive rates would help clarify the robustness of the proposed composition modeling.

A2:

Thank you for suggesting this test.

PARC can accept some unseen valid arrangements consistent with the variation in normal references, but this ability depends largely on the diversity of valid arrangements represented in those references. If the references cover the relevant local position variations and an unseen arrangement merely recombines these known states, it is more likely to be accepted. If reference arrangements are relatively uniform and an unseen valid arrangement substantially changes regional distributions or context, it is more likely to be misclassified as anomalous. The connector and breakfast-box experiments demonstrate these two cases.

Next, we first explain how the mean and covariance can allow the model to accept unseen valid arrangements, then use two experiments to examine the role of normal reference coverage. The connector experiment tests unseen combinations of known local positions. The breakfast-box experiment tests a valid fruit order absent from the fitting references. Both experiments report normal false-positive rates to assess how well PARC accepts unseen valid arrangements and how this depends on normal reference coverage.

Mechanism. The mean describes the typical composition. Covariance describes which statistics vary and which vary together. Positions that vary normally can receive greater tolerance, while departures from stable relationships receive larger Mahalanobis scores. These statistics and the alarm threshold jointly determine acceptance, so every valid combination need not be observed.

We would like to clarify that the distinction between valid arrangements and genuine logical anomalies depends on the inspection requirements. Without explicit rules, normal examples need to reflect the permitted variations.

Splicing connectors: an unseen combination of known local states. We test whether PARC accepts an unseen combination of familiar positions. References include each connector in higher/lower positions, but one complete left–right combination is excluded from the reference bank, statistical fitting, and calibration. In the covered control, that combination replaces 10% of the references without changing their total number. We repeat this over all four combinations, with five-fold evaluation keeping each source group intact.

Both conditions use the same 532 normal test images. Each sets its threshold from normal calibration data at nominal 5% FPR. FPR is the proportion of normal images incorrectly flagged. Results use complete PARC fusion.

Normal-reference conditionNormal false-positive rate ↓
Target valid position combination not covered6.18%
Target valid position combination covered3.90%

FPR increases by 2.28 percentage points. Each connector's local position is already represented, so the new combination can remain consistent with normal appearance and composition. The results show that PARC can accept some unseen valid combinations of these known local states.

Breakfast box: insufficiently represented fruit order. To test whether missing examples of a valid fruit order increase false alarms, we compare reference coverage. We fix reference count, final normal calibration images, test images, and the threshold-setting rule. Only coverage of the target valid fruit order in the fitting references changes. Each condition sets its threshold using the same normal calibration images.

Normal references used to fit the statisticsNormal false-positive rate ↓
Target valid fruit order not covered32.88%
Target valid fruit order covered3.62%

Coverage lowers FPR by 29.26 percentage points. Fruit reordering can change both regional distributions and contextual features despite unchanged types/counts. Departures along directions that previously varied little can receive high scores. The results show that covering this valid order reduces false alarms by improving the appearance references and composition statistics.

The distinction is known local states versus insufficiently represented distributions. Connector positions recombine previously observed states. Fruit reordering can introduce larger changes not covered by the original references. This explains the different outcomes and supports prioritizing permitted variations missing from normal references.

We have added these experiments to Appendix G (Table 15) (Page 20 of the revised manuscript).

Reviewer_iTn9OpenReview ↗
Rating
4(+1)
Confidence
3(−1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PARC, a training-free framework for jointly detecting structural and logical anomalies. It combines patch-level appearance matching with regional composition modeling, which captures the spatial distributions and proportions of recurring visual patterns in normal images. The two anomaly scores are calibrated and fused without additional network training.

Experiments on MVTec-LOCO-AD and three additional datasets demonstrate strong detection performance. Evaluations across five feature sources, together with component ablations and sensitivity analyses, support the effectiveness and efficiency of the approach.

Critical strengths
  1. Simple and effective approach to unified anomaly detection.

The paper addresses an important limitation of local appearance-based anomaly detection by introducing regional composition statistics without requiring additional neural-network training or heavy auxiliary modules. The method is computationally efficient and can be applied to different feature extractors.

  1. Comprehensive experimental validation.

The evaluation covers multiple datasets and feature sources, with detailed ablations demonstrating the complementary contributions of appearance and composition modeling. The sensitivity and efficiency analyses further strengthen the empirical findings.

Critical weaknesses

The proposed composition representation relies on regional proportions and global pattern statistics. While effective for anomalies involving component quantities and coarse spatial arrangements, it may be less suitable for more complex logical constraints, such as ordering, fine-grained spatial relationships, or interactions between individual components.

The current experiments do not sufficiently establish the boundaries of this representation. Consequently, the general applicability of PARC to different types of logical anomalies remains somewhat unclear.

Questions for authors

How well can PARC handle logical anomalies involving ordering or fine-grained relationships between components, where regional pattern proportions may remain nearly unchanged?

Could the authors further characterize which types of logical constraints can be effectively represented by regional composition statistics and which are likely to require more explicit relational modeling?

3 comments
AuthorsOpenReview ↗

Author Responses to Reviewer iTn9 (Part 1)

We would like to sincerely thank you for your efforts and valuable comments to improve our work!

Below we address your concerns.


Q1: Ordering, compatibility, and mismatched connection positions at the two cable ends.

How well can PARC handle logical anomalies involving ordering or fine-grained relationships between components, where regional pattern proportions may remain nearly unchanged?

A1(part 1):

Thank you for asking us to clarify these capabilities.

PARC detects some fixed-order and compatibility errors, but remains limited on unseen configurations involving multiple valid orders and mismatched connection positions at the two cable ends. The key is whether the relationship changes survive feature extraction and aggregation.

To explain why PARC handles these cases differently, we first examine which relationship information its representation keeps. We then test order changes with nearly unchanged regional proportions. This separates an unusual part position from an invalid order made up of familiar parts and positions. Next, connector mismatches test whether cross-block covariance detects normal parts paired incorrectly. Finally, we test mismatched connection positions at the two cable ends to assess detection of specific connection errors.

Mechanism. The composition branch first aggregates regional patterns into a descriptor, then assesses whether it deviates from the normal distribution. Regional statistics [Eqs. (4)–(5)] average patches' responses to normal visual prototypes to obtain the descriptor hh. This aggregation retains each region's pattern composition but does not directly record the order within a region.

Given this descriptor, the raw composition score [Eq. (6)] is

L(X)=(h(X)−μh)⊤Σh−1(h(X)−μh),L(X)=(h(X)-\mu_h)^\top\Sigma_h^{-1}(h(X)-\mu_h),

where μh\mu_h and Σh\Sigma_h are the mean and covariance of normal descriptors. Full covariance uses relationships across regions and between regional and global statistics to identify abnormal combinations, but cannot recover ordering information already lost during aggregation.

Therefore, swapping only the positions of already extracted features within the same region leaves both the descriptor and composition score unchanged. When parts are rearranged in the original image, however, changes in position and context may alter the newly extracted features and their prototype responses. This can change the descriptor, allowing the composition branch to detect some ordering errors.

Ordering experiment. To test whether PARC detects order changes despite nearly unchanged regional proportions, we rearrange the same connector crops A/B/C across three fixed slots within one coarse region, preserving counts, background, and occupied slots, then re-extract features. Each source group contains three connector crops, A/B/C, taken from three different original photographs. All arrangements generated from these crops belong to the same group and stay in the same data split. Original photographs are not reused across groups. Across the 60 test groups, the mean total-variation distance (a measure of their difference) between regional proportions is 0.00526.

With only ABC accepted, ACB moves B/C away from their normal positions. With ABC/BCA/CAB accepted, every part has appeared at every slot. The other three permutations use those same parts and slots. Detecting them therefore requires checking combinations across positions, rather than each position independently. The first row averages all six choices of one accepted order. The second averages two complementary sets of three. Values are AUROC (%), averaged over the corresponding rules. Branch scores are raw, while fusion uses normal calibration.

Accepted ordersPatch appearance branchRegional composition branchDefault fusion
One accepted order, averaged over all six choices99.9181.3797.86
Three accepted orders, averaged over both complementary groups59.6456.9859.46

The results show that PARC can use fixed-position cues, while invalid combinations of familiar parts and positions are harder to detect. The strictness of the composition branch also depends on how broadly normal orderings are covered. In general, allowing fewer normal orderings yields a smaller descriptor variance, so deviations from them receive stronger penalties. Allowing more normal orderings broadens the normal distribution and increases its variance, making the model more tolerant of these changes rather than treating them directly as defects.

AuthorsOpenReview ↗

Author Responses to Reviewer iTn9 (Part 2)

A1(part 2)

Compatibility experiment. To test whether cross-block covariance detects wrong pairings of normal parts, we use splicing connectors. In this category, three clamps on each side or five on each side is normal. Three on one side and five on the other is abnormal. Each clamp count is normal independently, so the combination matters. We evaluate all 9 such anomalies and 119 normal images, comparing PARC without cross-block covariance and full PARC.

SettingPatch appearance branchRegional composition branchDefault fusion
PARC (without cross-block covariance)68.4774.4870.96
Full PARC68.4790.6085.15

Holding the descriptor and within-block covariance fixed, restoring cross-block relationships raises composition-branch AUROC from 74.48 to 90.60. The result shows that cross-block relationships help detect wrong pairings of normal connector types.

Fine-grained relationship experiment. To test PARC's ability to detect logical anomalies involving fine-grained relationships, we use mismatched connection positions at the two cable ends as an example. We evaluate all 25 images of this type. Composition AUROC is 60.61. The cable connects at different relative positions at its two ends, while counts, types, and coarse distributions may stay similar. PARC does not directly record which cable connects to which clamp. Finer spatial or explicit relationship representations are needed when aggregation loses this evidence. The result shows that the current statistics are less effective at detecting mismatched connection positions at the two cable ends.

We have clarified the mechanism in Section 3.3 (Page 5 of the revised manuscript) and added these tests to Section 4.2 (Table 6) (Page 10 of the revised manuscript) and Appendices F.2–F.3 (Page 19–20 of the revised manuscript).

AuthorsOpenReview ↗

Author Responses to Reviewer iTn9 (Part 3)

Q2: Strengths and boundaries of the composition representation.

Could the authors further characterize which types of logical constraints can be effectively represented by regional composition statistics and which are likely to require more explicit relational modeling?

A2:

Thank you for asking us to identify the boundary.

PARC is strong on missing parts, overall quantity changes, clear layout errors, and some content–label mismatches. Its main difficulties are specific part relationships, exact local counts and positions, and precise dimensions. To identify which logical constraints PARC handles well, we compare defect types using Logical AUROC from default complete PARC. All values below are Logical AUROC (%), measuring the model's ability to distinguish normal images from the corresponding anomalies.

Regional composition statistics are well suited to representing constraints on component presence, overall quantities and proportions, regional configurations, and some content–label matches. Component presence, quantity, and proportion constraints can be reflected by whether the corresponding visual patterns appear and whether their proportions are normal. Regional configuration constraints can be reflected by the regions in which these patterns occur. Content–label matching constraints can be reflected by normal relationships between their statistics. In the corresponding experiments, assembly-component quantity errors, missing required labels, and missing bottle liquid each achieve 100.00% Logical AUROC. Missing ingredients and incorrect ingredient proportions achieve 96.47% and 99.75%, respectively. For regional configuration and content matching, swapped upper and lower labels, ingredients placed in the wrong compartments, and juice–label mismatches also each achieve 100.00%.

The weaker results identify three specific boundaries:

Main limitationDefect subtypeAnomalous imagesComplete PARC Logical AUROC (%)
Specific part relationshipsMismatched connection positions at the two cable ends2560.53
Local counts and positionsIncorrect pushpin count within compartments7874.95
Exact dimensionsIncorrect screw length6578.86
Specific part relationshipsCable color inconsistent with connector clamp count1284.19
Local counts and positionsFruit icon in the wrong position1084.88

Specific part relationships: regional statistics describe which patterns occur, not which cable connects to which clamp. Changes that preserve counts, types, and regional distributions can leave little evidence, and global matching may still explain each local patch. Cable connection positions must correspond at both ends. Cable color must match connector clamp count. AUROC is 60.53% and 84.19%, respectively, supporting the need for explicit association or connectivity information.

Exact local counts and positions: regional averaging can mix missing and extra items, while small displacements may leave region membership unchanged. Pattern proportions are not instance counts, and whole-image totals cannot recover lost within-region locations. Per-compartment pushpin counts reach 74.95% and misplaced fruit icons 84.88%, versus 100% for overall connector-component counts and fruit/“100% Juice” label swaps. Clear overall changes leave stronger evidence than fine local allocation.

Exact dimensions: PARC responds to the visual effects of length changes without measuring endpoints or scale directly. Screw-length errors reach 78.86%, whereas cable-length errors reach 100%: effectiveness depends on how an object presents its dimensional differences. When necessary details are lost, finer spatial representations, instance counting, geometric measurement, or explicit relationships can provide the missing evidence.

We have added this analysis to Appendix F.1 (Table 12) (Page 17–19 of the revised manuscript) and clarified the scope of PARC in Section 4.2 (Page 10 of the revised manuscript).

Reviewer_LVMCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PARC, a training-free method for detecting structural and logical anomalies in industrial images. It has two branches. The appearance branch uses nearest-neighbor patch matching to detect local defects. The composition branch models the distribution of visual patterns across image regions. PARC then calibrates and combines the two scores for anomaly detection.

Critical strengths
  1. Patch-based methods can miss logical anomalies because individual components may look normal even when they are misplaced. Regional composition modeling is a reasonable way to address this issue.

  2. The method is simple and requires no category-specific network training or additional segmentation models. It can also be applied to different feature extractors.

  3. The results are promising, especially for logical anomaly detection. The average combined AUROC across five feature extractors increases from 72.69% to 84.74%. The ablation studies are useful, and the additional inference cost is low.

Critical weaknesses
  1. The novelty is somewhat limited. K-means clustering, regional histograms, Mahalanobis distance, and score calibration are well-established techniques. The paper mainly combines these methods. In Table 5, the full covariance model improves combined AUROC by only 0.55 percentage points over the block-diagonal model. The main contribution of joint modeling remains unclear.

  2. The maximum-based score fusion does not always improve performance. In Table 4, the composition branch achieves 90.88% logical AUROC, but fusion reduces it to 89.42%. Similar decreases occur in four of the five categories in Table 9.

  3. The improvement over LogSAD is small. PARC achieves 91.12% combined AUROC, compared with 90.24% for LogSAD. CSAD and LogiCo also achieve higher accuracy. PARC offers efficiency and training-free advantages, but its accuracy gains remain modest.

  4. The evaluation of logical anomaly detection mainly relies on MVTec-LOCO-AD, which contains only five categories. The fixed spatial regions may also be sensitive to object shifts, rotations, and complex component arrangements. These cases are not fully evaluated.

Questions for authors
  1. What is the main novelty of PARC compared with ComAD, CSAD, and SINBAD? What benefits does full covariance modeling provide, given the small improvement in Table 5?

  2. Why does score fusion reduce logical anomaly detection performance in Table 4? Have the authors tested other strategies, such as weighted or adaptive fusion?

  3. PARC improves combined AUROC over LogSAD by only 0.88 percentage points. Could the authors provide a controlled comparison using the same feature extractor and normal training data?

  4. How robust is PARC to object shifts and rotations? Could the authors test more diverse logical anomalies to better evaluate its generalization ability?

4 comments
AuthorsOpenReview ↗

Author Responses to Reviewer LVMC (Part 1)

We would like to sincerely thank you for your efforts and valuable comments to improve our work!

Below we address your concerns.


Q1a: Method contribution.

What is the main novelty of PARC compared with ComAD, CSAD, and SINBAD?

A1a:

Thank you for asking us to clarify the contribution.

PARC extends local appearance detection with a composition check on the same frozen features, without component masks, neural-network training for each product category, or explicit product rules.

Differences from existing methods.

  1. ComAD extracts component regions and measures their areas, colors, and counts. Boundary or instance-separation errors may affect subsequent decisions. PARC directly aggregates visual pattern responses without forming complete component regions, reducing dependence on segmentation quality and region postprocessing costs.

  2. CSAD generates component pseudo-labels and trains a category-specific segmentation network, requiring additional preparation and training. PARC builds normal composition statistics directly from frozen features, avoiding component pseudo-labels and category-specific network training and reducing adaptation costs for new products.

  3. SINBAD represents feature sets with random-projection histograms and uses nearest-neighbor scoring. Discrete bins ignore within-bin variation, and scoring requires retrieving normal descriptors. PARC jointly describes regions using continuous responses to normal prototypes, avoiding hard binning and nearest-neighbor retrieval of composition descriptors.

MethodExisting designPARC differencePARC advantage
ComADComponent regions with area, color and count statisticsRegional visual-pattern responses without explicit component regionsLess dependence on component boundaries and instance separation, reduced postprocessing cost
CSADComponent pseudo-labels and category-specific segmentation trainingNormal statistics fitted directly to frozen featuresNo pseudo-label preparation or category-specific network training, lower adaptation cost
SINBADRandom-projection histograms and nearest-neighbor descriptor scoringContinuous prototype responses, joint regional statistics and Mahalanobis scoringNo hard binning or nearest-neighbor retrieval of composition descriptors

We compare five feature sources to assess performance and added runtime. Composition scoring and fusion add 1.83 ms/image. Mean Combined AUROC rises from 72.69 to 84.74, without preparing component labels or training a segmentation network for each source. Composition improves average detection at little added cost.

We have clarified PARC's contributions in the Introduction (Page 2–3 of the revised manuscript).


Q1b: Benefits of full covariance.

What benefits does full covariance modeling provide, given the small improvement in Table 5?

A1b:

Thank you for asking us to clarify the benefits of full covariance.

Full covariance mainly provides additional evidence for individually normal parts combined incorrectly. Its Mahalanobis score penalizes deviations along directions with little normal variation. Thus, individually plausible statistics can still form an unusual combination. A block groups statistics for one region or the whole image. Cross-block covariance captures relationships between regions and between regional and whole-image statistics. Block-diagonal covariance keeps only relationships within each block.

To test whether cross-block covariance helps detect wrong combinations of normal parts, we use splicing connectors as a representative compatibility test. Both sides having three cable clamps or both having five is normal. Three on one side and five on the other is abnormal. We evaluate all 9 such anomalies and 119 normal test images. The rows compare whole-image statistics and regional statistics without and with cross-block covariance.

Composition representation and covariance settingRegional composition AUROC (%)
PARC (no partitioning)73.39
PARC (partitioned, without cross-block covariance)74.48
Full PARC (partitioned, with cross-block covariance)90.60

Holding the regional descriptor and within-block covariance fixed, cross-block covariance raises AUROC from 74.48 to 90.60. Whole-image statistics and regional block-diagonal modeling perform similarly. Cross-block covariance helps detect these mismatches between normal part types, while its average fused Combined AUROC gain remains 0.55 points. The composition representation is the main contribution. Full covariance adds benefits for specific compatibility constraints.

We have clarified the role of full covariance in Section 3.3 (Page 5 of the revised manuscript) and added the compatibility test to Section 4.2 (Table 6) (Page 10 of the revised manuscript).

AuthorsOpenReview ↗

Author Responses to Reviewer LVMC (Part 2)

Q2a: Why fusion lowers Logical AUROC.

Why does score fusion reduce logical anomaly detection performance in Table 4?

A2a:

Thank you for examining this tradeoff.

Max can retain a high appearance score for a normal image, placing it above a logical anomaly that composition alone ranked correctly. A product may have correct quantities and combinations but differ visually from its references. Its normal-image score then rises, while a logical anomaly with normal-looking parts may receive little increase. AUROC measures their ordering, so it can fall even though max lowers no individual score. To test whether ranking changes explain the drop, we compare normal–anomaly pairs before and after fusion. Pairwise analysis attributes 3.37 points of loss to worsened rankings and 1.91 points of gain to improved rankings: 90.88 → 89.42, a 1.46-point decrease. The result shows that ranking losses outweigh gains, explaining the lower Logical AUROC.

The two branches specialize in different defects, so fusion balances the two anomaly types and can sacrifice performance on one anomaly type. Most tested rules show this tradeoff for logical anomalies. For structural anomalies, however, max exceeds either branch: stains or missing material can also change regional pattern proportions and whole-image totals, allowing composition to catch defects missed by appearance. We repeat the pair analysis for structural anomalies to measure the benefit of adding composition. Worsened rankings cost approximately 1.40 points, improved rankings gain 2.39, and the net gain is approximately 0.98 points, corresponding to 92.60 → 93.59. The result shows that composition adds useful structural evidence whose ranking gains outweigh the losses.

We have clarified the fusion tradeoff in Section 4.2 (Page 8–9 of the revised manuscript) and added the pairwise analysis to Appendix H.2 (Page 21 of the revised manuscript).


Q2b: Weighted and adaptive fusion.

Have the authors tested other strategies, such as weighted or adaptive fusion?

A2b:

Thank you for suggesting alternative fusion strategies.

To test whether other fusion rules improve the balance between anomaly types, we compare averaging, 109 fixed weights, and one image-dependent rule at six temperatures using the same calibrated branch scores. The image-dependent rule is:

w_Q(x) &= \mathrm{sigmoid}\left(\frac{u_Q(x)-u_L(x)}{T}\right), \\\\ S(x) &= w_Q(x)u_Q(x) + (1 - w_Q(x))u_L(x). \end{aligned}$$ Here, $u_Q,u_L$ are calibrated patch appearance and regional composition scores. $Q,L$ denote raw scores in the paper. The sigmoid converts the difference to a weight between zero and one. We test $T \in \lbrace 0.001, 0.005, 0.01, 0.05, 0.1, 0.5 \rbrace$: smaller $T$ approaches max. Larger $T$ approaches equal averaging. Each setting fixes $T$, while weights vary by image. Results are averaged over five categories. In the table, “Patch appearance only” and “Regional composition only” use their respective calibrated branch scores. “Equal-weight average” takes their mean. “Best tested fixed weighting” uses the same weight for all images. “Best tested image-dependent weighting” adjusts the weight for each image using the formula above, with $T=0.001$. “Max, default” takes the higher score. Both “Best tested” settings are chosen by test Combined AUROC. | Scoring rule | Combined AUROC | Structural AUROC | Logical AUROC | |---|---:|---:|---:| | Patch appearance only | 81.93 | 92.60 | 73.58 | | Regional composition only | 87.42 | 83.82 | 90.88 | | Equal-weight average | 85.69 | 89.18 | 83.36 | | Best tested fixed weighting* | 87.64 | 84.40 | 90.81 | | Best tested image-dependent weighting* ($T=0.001$) | 91.13 | 93.65 | 89.38 | | Max, default | 91.12 | 93.59 | 89.42 | \* “Best” is selected afterward by test Combined AUROC to analyze the tested rules, not to choose deployment parameters. The best image-dependent setting, $T=0.001$, already approximates max: Combined AUROC is 91.13 versus 91.12. **The results show that max performs nearly as well as the best tested weighting without choosing a weight or temperature.** Relative to composition alone, it gains 3.70 Combined points at a cost of 1.46 Logical points. Composition alone is preferable for logical-only evaluation here. We have added the weighting comparison to Appendix H.2 (Table 18) (Page 21–22 of the revised manuscript).
AuthorsOpenReview ↗

Author Responses to Reviewer LVMC (Part 3)

Q3: Controlled comparison with LogSAD.

PARC improves combined AUROC over LogSAD by only 0.88 percentage points. Could the authors provide a controlled comparison using the same feature extractor and normal training data?

A3:

Thank you for suggesting a controlled comparison.

With the same features and normal training data, PARC outperforms LogSAD with all three backbones. Explicit rules give LogSAD larger gains and narrow the gap.

To remove differences in feature extractors and measure the extra benefit of explicit rules, we compare the methods with three backbones. Each comparison uses the same backbone weights and normal training images. “Before rules” uses no explicit rules. “After rules” applies the same LogSAD rule alarms to the final scores. These rules impose specific constraints, such as requiring exactly 15 pushpins and agreement between juice color and the fruit label. This extra semantic judgment is borrowed from LogSAD, not learned by PARC's composition representation. The table reports Logical AUROC (%).

Scoring backboneLogSAD: before → after rulesPARC: before → after rules
RADIO (AnomalyVFM weights)81.53 → 90.2589.13 → 91.63
DINOv277.40 → 88.9787.33 → 90.18
CLIP80.42 → 89.3285.73 → 89.41

The results show that PARC leads by 5.31–9.93 Logical AUROC points without rules. Shared rules add 8.71–11.57 points to LogSAD and 2.50–3.68 points to PARC. Even after adding the extra semantic judgments designed for LogSAD, PARC retains a small performance advantage. Explicit rules therefore provide an important part of LogSAD's final performance.

Product-specific rules require design, implementation, validation, and maintenance. Changes to required counts or compatibility relationships require rule updates, limiting transfer to new products and defect types. Default PARC instead estimates composition statistics directly from normal images and avoids both rule preparation and component segmentation with the Segment Anything Model (SAM).

The default complete system is also faster: To compare complete-system accuracy and runtime, Table 1 reports 91.12 Combined AUROC at 114.35 ms/image for PARC versus 90.24 at 2138.66 ms/image for LogSAD, whose total runtime is 18.7 times higher. The results show that default PARC combines higher Combined AUROC with much shorter runtime.

AuthorsOpenReview ↗

Author Responses to Reviewer LVMC (Part 4)

Q4a: Robustness to shifts and rotations.

How robust is PARC to object shifts and rotations?

A4a:

Thank you for asking about robustness to shifts and rotations.

Normal-reference augmentation substantially reduces PARC's degradation under the tested shifts and rotations.

Reference augmentation mitigates placement errors. To test whether normal movement examples improve detection after shifts and rotations, we expand the normal references. Each normal training image contributes seven views: the original, center rotations of ±15°, horizontal translations of ±15% of image width, and vertical translations of ±15% of image height. All views expand both the appearance reference bank and composition-fitting data, without additional network training.

We test horizontal and vertical translations separately at 15±515\pm5% of width and height, and rotation magnitudes of 15∘±5∘15^\circ\pm5^\circ. Tests use the same augmentation configuration, with results averaged over five categories. Each model's alarm threshold stays fixed across test transformations. FPR is the percentage of normal images incorrectly flagged.

Horizontal and vertical translations: We test the directions separately to check whether the effect depends on movement direction.

Test transformationOriginal Combined AUROCAugmented Combined AUROCOriginal FPRAugmented FPR
Horizontal translation, magnitude 15±515\pm5%76.9291.9579.3614.94
Vertical translation, magnitude 15±515\pm5%74.5591.3898.9915.95

Translations change object positions in the image and may alter visual pattern proportions within fixed regions, making the original configuration sensitive to translated test images. After reference augmentation, AUROC reaches 91.95 for horizontal translations and 91.38 for vertical translations, with FPR falling to around 15% in both cases. Adding normal translation examples to the reference helps PARC avoid treating these placement changes as defects.

Rotations: We test if the same reference expansion also improves detection after orientation changes.

Test transformationOriginal Combined AUROCAugmented Combined AUROCOriginal FPRAugmented FPR
Rotation, magnitude 15∘±5∘15^\circ\pm5^\circ75.9291.1386.0313.45

Augmentation raises rotation AUROC from 75.92 to 91.13 and lowers FPR from 86.03 to 13.45. The same reference expansion also improves detection after rotation.

However, on a production line, unexpected product translations or rotations may indicate positioning or conveyor problems, so the original configuration's sensitivity to these changes can also be useful. If these changes are acceptable placement variations, augmenting normal reference images and expanding both the appearance reference bank and composition statistics can substantially reduce the resulting false alarms, without changing the model architecture or retraining the feature extractor.

We have added the reference augmentation results to Section 4.2 (Page 10 of the revised manuscript) and Appendix H.4 (Table 20) (Page 22 of the revised manuscript).


Q4b: Broader logical anomaly evaluation.

Could the authors test more diverse logical anomalies to better evaluate its generalization ability?

A4b:

Thank you for suggesting broader evaluation.

PARC leads the tested methods on three additional evaluations.

Broader logical inspection. To test performance on more products and inspection tasks, we add VID-AD [1] and MIAD [2] and extend the baseline comparison on CAD-SD [3]. VID-AD covers all 50 tasks: ten scenarios under five capture conditions, checking quantity, type, length, placement, and relationships. MIAD (Maintenance Inspection Anomaly Detection) covers its three logical categories: Catenary Dropper, Nut and Bolt, and Witness Mark—overhead-wire suspension parts, fasteners, and alignment marks. CAD-SD (Co-occurrence Anomaly Detection Screw Dataset), already evaluated in the paper, includes both local defects and assembly errors such as extra or missing nuts. We report image-level AUROC.

MethodVID-ADMIADCAD-SD
PatchCore50.2373.5281.90
AnomalyDINO68.2867.0690.50
SubspaceAD60.0061.2285.74
PARC75.9475.6999.81

PARC outperforms the tested baselines on these additional products and inspection tasks beyond MVTec LOCO AD (LOCO).

We have added these evaluations to Appendix I (Table 21) (Page 22–23 of the revised manuscript).

[1] Hiroto Nakata, et al. VID-AD: A Dataset for Image-Level Logical Anomaly Detection under Vision-Induced Distraction, arXiv:2603.13964, 2026.

[2] Tianpeng Bao, et al. MIAD: A Maintenance Inspection Dataset for Unsupervised Anomaly Detection, arXiv:2211.13968, 2022.

[3] Kengo Ishida, et al. SA-PatchCore: Anomaly Detection in Dataset With Co-Occurrence Relationships Using Self-Attention, IEEE Access, 2023.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.