acceptodds
Under review as a conference paper at ICLR 2027

Match Every Patch Once: A Logical-Structural Tradeoff in Anomaly Scoring

Abstract

Normal-only image anomaly detection often scores an image by matching each test patch independently to its closest patch in a normal reference. This allows many test patches to match the same reference patch, making changes in the number or composition of otherwise normal parts difficult to detect. We instead require a one-to-one patch assignment between a test image and a complete normal reference, keeping the features, patch costs and aggregation fixed. We define the assignment excess as the additional cost introduced by one-to-one matching and show that it exactly accounts for the resulting changes in anomaly ranking. The constraint improves sensitivity to logical anomalies but can reduce sensitivity to structural defects, a tradeoff that depends strongly on how normal references are constructed. Used as an auxiliary signal in a trained EfficientAD-M detector, a calibrated form of the assignment excess improves logical anomaly detection in all five training runs. Patch correspondence and reference construction must therefore be specified together in anomaly scoring.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
anomaly detection, logical anomaly detection, industrial anomaly detection, optimal transport, patch matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GRp5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the effect of replacing reusable nearest-neighbor patch matching with one-to-one assignment in normal-only image anomaly detection. The authors define assignment excess as the additional cost introduced by the one-to-one constraint and show that it accounts for changes in anomaly ranking. On MVTec LOCO, one-to-one matching consistently improves logical anomaly AUROC while reducing structural anomaly AUROC, with the gain mainly arising from count and composition anomalies rather than pure rearrangements. The paper further investigates reference construction and shows that pooling multiple normal references substantially weakens the effect, whereas scoring complete references independently largely preserves it. Additional experiments on CAD-SD, DigitAnatomy, VID-AD, VisA, and Real-IAD, together with integrations into PatchCore and EfficientAD-M, are used to examine the generality and practical utility of the proposed matching signal.

Critical strengths
  1. The paper provides a clean controlled study of the correspondence constraint itself. The nearest-neighbor and one-to-one variants share the same feature representation, normal reference, patch costs, and aggregation, so the observed performance difference can be attributed much more directly to whether reference patches may be reused. The assignment-excess formulation further gives a simple explanation of how this constraint changes normal-anomaly rankings, rather than treating the improvement as a purely empirical effect.

  2. The analysis reveals a useful interaction between correspondence and reference construction. The logical gain is consistently observed with complete single references, but becomes much smaller when multiple references are pooled and can be recovered when complete references are scored separately. Together with the consistent logical-structural tradeoff across DINOv2 and DINOv3, this provides a useful characterization of when one-to-one matching is likely to help and when it may fail.

Critical weaknesses
  1. The methodological novelty appears limited relative to recent logical anomaly detection work. One-to-one, Hungarian, and transport-based correspondence have already been used in related methods, while the main technical contribution here is to isolate this matching constraint and analyze its effect under controlled conditions. The assignment-excess identity is useful for interpretation, but it is essentially the residual between the assignment and nearest-neighbor objectives.

  2. The claimed applicability to logical anomaly detection is substantially narrower than the terminology suggests. The proposed score depends only on the multiset of patch features and is therefore insensitive to rearrangements that preserve this multiset; correspondingly, the paper reports essentially no gain for pure rearrangement anomalies. In addition, the method relies on a complete normal reference from the correct category, while the experiments show that pooling references can largely remove the logical advantage. These properties raise concerns about whether the method addresses logical anomalies broadly, or mainly a restricted subset involving count and composition changes under favorable reference construction.

Questions for authors
  1. Since related logical anomaly detection methods already employ Hungarian or transport-based matching, can the authors clarify more concretely what capability is introduced by the proposed method beyond providing a controlled analysis of an existing correspondence mechanism? A comparison under the same backbone, reference construction, and scoring protocol with a recent matching-based logical anomaly detector would help establish the incremental methodological contribution.

  2. How does the method perform when reference retrieval is not restricted by the ground-truth category? The current LOCO setup retrieves a normal reference within the correct category, but this information may not be available in a more general anomaly detection setting. A category-agnostic retrieval experiment or a controlled routing-error experiment would help quantify this dependency.

  3. The method is invariant to rearrangements that preserve the patch multiset, and the reported placement-only results show almost no improvement. How should the method be expected to handle logical anomalies defined primarily by spatial relationships or ordering rather than count or composition? Would the authors consider narrowing the scope of the main claim accordingly?

  4. The paper consistently observes a gain on logical anomalies together with a loss on structural anomalies at the matching-score level. Is there a principled way to combine the reusable and one-to-one scores so that the logical gain can be retained without sacrificing structural performance, rather than requiring the user to choose between the two behaviors?

4 comments

Hungarian matching inside the related methods you mention (LogSAD, ObjectCore) is known; what we add is that, on the same features and costs, the one-to-one constraint helps when each reference stays complete and gives almost nothing on a pooled bank: +5.69 [+4.20, +7.22] logical AUROC on separate references against +0.24 [-0.04, +0.55] pooled, on DINOv2 (Table 10a).

Q1. In Table 10a, the bank decides whether one-to-one matching helps. The pooled difference is too close to call. We change only the assignment and the bank, so this isolates what one-to-one matching adds, the interaction you call "a useful interaction between correspondence and reference construction". Inside the first of these, as the official four-shot system with its CLIP features, supports, calibration and fusion kept, our assignment beats reuse on the same dense costs by +5.27 [1.73, 8.81] branch logical AUROC, positive in all five categories. Replacing its Segment Anything region branch with this mask-free dense assignment gives 85.80 against 87.01 fused logical AUROC (-5.22 to +2.82), too close to call. After fusion the branch difference is 2.23 [-2.75, 7.21], which includes zero (Appendix H, "Assignment inside LogSAD"). That test is four-shot, on development data, and a pre-specified arm. In practice this gives a rule for patch-level one-to-one scoring, which is to keep the normal references complete and score them separately. Among the systems in Table 19, ours is also the only one that forbids patch reuse with no segmenter or detector, and the controlled result in Table 10a needs no trained component. For detection in practice we add it to a trained host detector, here EfficientAD-M (Q4).

On the support images of the four-shot system, a recent matching-based detector (four normal reference images per category), DINOv2 patch assignment with one retrieved reference gains +5.16 over nearest neighbor for one support draw (Appendix H). It matches Segment Anything regions described with CLIP. On the 606 of 1,568 queries with a single region, its Hungarian step and reuse score identically. With calibration fixed, Hungarian minus reuse in the four-shot system is +0.01 logical AUROC over five categories, too close to call (Appendix H, "Matching within LogSAD").

Inside the system's region-level step, with the same support images, the interaction between pooled versus separate references and Hungarian matching versus reuse is -0.45 logical AUROC over five categories, too close to call. The four-shot system's own patch branch allows reuse on a pooled bank (Table 19), so our rule would apply there only if one-to-one matching were added. Our bank rule is about patch-level one-to-one matching, which our main contrasts (Tables 1 and 10a) control; region-level matching and full systems lie outside it.

Q2. Category-agnostic retrieval picks the same reference as correct-category retrieval for all 1,568 test images. Here we use DINOv2 retrieval without category labels (Appendix H, "Category routing"). When we force 10% wrong-category errors, assignment-score logical AUROC drops from 78.30 to 73.07, and structural AUROC also falls.

Q3. Yes: our claim now covers count and composition changes. Our contributions in Section 1 now say so. Count and composition are where reuse-based scoring is weakest. With equal patch counts and discrete patch types shared with the reference, Proposition 1 shows that reuse misses a count change, while the assignment cost is bounded below in proportion to the difference in type counts. A rearrangement that preserves the patch-feature multiset leaves the assignment cost unchanged, and our exploratory LOCO split in Appendix H agrees. Of the 561 logical anomalies, 502 are outside the group that only rearranges parts, and assignment gains +5.29 over nearest neighbor on them. On the 59 placement-only anomalies, in three categories, the difference is -0.38, and its interval spans zero.

Q4. Yes: added to a trained host's score, the conditional excess raises logical AUROC. On EfficientAD-M we keep the host's score, calibration and fusion weight fixed and add one statistic at unit weight. That statistic is the excess minus the median excess of the nearest fitting normals, scaled without anomaly labels, keeping only its positive part (Appendix H, "Score fusion"). Across three 70,000-update seeds, this raises logical AUROC from 82.51 for the host alone to 88.08, while structural AUROC moves from 89.81 to 90.10, against 90.23 with Chamfer fusion. The gain over Chamfer is +2.55 (95% range over seeds 1.38 to 3.72; Appendix H, "Longer host"). On PatchCore the signal also lifts logical AUROC, at a structural cost of 1.4 points, beyond our harm margin, so EfficientAD-M is the host that keeps both.

A pooled bank is the default in the patch detectors of Table 19, including the four-shot system's patch branch, so a null for one-to-one matching on such a bank is not evidence against the constraint on complete references.

I thank the authors for their response and additional experiments. My main concern about methodological novelty remains. The new experiments help explain when one-to-one matching is effective, but the main contribution is still the analysis of an existing matching mechanism. The fusion results on PatchCore also show that the trade-off between logical and structural anomaly detection remains an issue. I maintain my score of 2.

Thank you for the follow-up, and for calling this a "clean controlled study".

Systems such as LogSAD use one-to-one matching inside larger pipelines, so their papers do not show what the matching itself does. The paper takes it out and tests it alone. Forbidding reuse lets a patch detector notice a missing or extra part when each normal image is kept whole and not pooled into one bank.

The tradeoff is expected, because the two scores catch different defects. Added to a trained EfficientAD-M, the one-to-one signal catches more logical anomalies, and structural detection stays essentially the same.

AuthorsOpenReview ↗

Follow-up (18 Nov): the finding applied to an existing detector

Thank you for your comment. This follows up our 17 Nov reply with a new result. By design, the paper takes the one-to-one constraint out of larger systems and tests it alone. Its practical finding is to keep each reference whole, and the table adds our calibrated one-to-one signal, computed against complete references, to our AnomalyDINO reimplementation (exploratory, from cached features).

AnomalyDINO reimplementation plus our signalLogical change, pointsStructural change, points
DINOv2+5.30 [3.96, 6.67]-0.22 [-0.94, 0.45]
DINOv3+4.38 [2.89, 5.94]-1.79 [-2.70, -0.91]

Under DINOv3 the structural cost exceeds our one-point margin. For DINOv2 the structural-change interval spans zero, and on the released EfficientAD-M checkpoint structural AUROC moves from 95.19 to 95.05. We would welcome your assessment of these results.

Reviewer_752bOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

This paper investigates the impact of patch matching rules on anomaly detection. Under the conditions of fixed features, normal references, cost matrices, and mean aggregation, the authors replace the nearest neighbor scoring, which allows reference reuse, with one-to-one matching, and define the additional matching cost as assignment excess. The core perspective is that prohibiting reuse can improve sensitivity to quantity and compositional anomalies, but may harm structural anomaly detection, and this trade-off depends on how the normal references are constructed. The authors further integrate the excess into EfficientAD-M and PatchCore to evaluate its role as an auxiliary detection signal.

Critical strengths

s1: The research problem has practical significance. Local appearances being normal but with incorrect quantities or compositions is a real challenge in industrial anomaly detection. Comparing different matching rules on the same cost matrix helps to understand the impact of correspondence constraints.

s2: The experimental analysis is relatively comprehensive. The authors investigate different encoders, reference constructions, and datasets, providing paired statistics and five-seed fusion results. Meanwhile, the paper reports the decline in structural performance and the reversal of benefits under certain settings, which helps in understanding the applicable scope of the method.

Critical weaknesses

w1: The core novelty of this paper remains unclear. One-to-one matching is sensitive to quantity and compositional changes, and existing matching methods have also utilized similar properties. The rank decomposition of the Excess mainly comes from its definition. The authors should explicitly clarify what new mechanistic insights this paper provides relative to existing work.

w2: The comparison with representative methods is insufficient. Although methods like ObjectCore and LogSAD are discussed, there is a lack of performance comparison, and RAD [1] is not included in the comparison either. The authors should explain under comparable settings what practical benefits the one-to-one constraint brings relative to existing logical anomaly and retrieval methods.

w3: The experiments in Appendix A.3 fail to independently distinguish the roles of coverage and prohibited reuse. Coverage-only changes the matching direction, and capacity relaxation simultaneously alters the coverage requirements. Therefore, attributing approximately half of the gains to the independent contribution of coverage still lacks sufficient evidence.

w4: The primary fusion benefits are achieved on a relatively weak EfficientAD-M baseline model, whose performance is significantly lower than the original paper's results listed in the appendix. The five-seed experiments support the improvements in the current configuration, but cannot rule out the possibility that the auxiliary branch is primarily compensating for insufficient training of the baseline model. The authors should verify the gains relative to the NN and Chamfer branches on a fully trained EfficientAD-M.

[1] Xingwu Zhang, Guanxuan Li, Paul Henderson, Gerardo Aragon-Camarasa, Zijun Long. “Is Task-Specific Training Necessary for Anomaly Detection?” ICML 2026.

Questions for authors

See the weaknesses.

5 comments

You ask (w1) "what new mechanistic insights this paper provides relative to existing work". ObjectCore already matches against a complete support image, and LogSAD already uses Hungarian matching; what we add is that, on the same references, features and costs, the patch-level gain of assignment over nearest neighbor nearly disappears when the references are pooled and returns when each complete reference is scored separately: +5.69 [+4.20, +7.22] logical AUROC on separate references against +0.24 [-0.04, +0.55] pooled, on DINOv2 (Table 10a). With five references, the pooled interval is too close to call. So patch-level one-to-one matching helps on complete references and does almost nothing on a pooled bank of five references, the kind of bank the patch detectors reviewed in Table 19, including LogSAD's patch branch, store. Our rule for a patch-level one-to-one detector is to keep the references complete and score them separately.

w1: The bank decides whether one-to-one matching helps, and Section 3 gives the mechanism for one complete reference. Proposition 1 is not the identity. Under discrete types, an idealized case, it shows that every nearest-neighbor score is zero, while the assignment cost is bounded below by the difference in type counts. A pooled bank breaks that premise, since it can supply missing counts from other images, and Table 10a measures the result. Proposition 2 gives a continuous substitution case where one-to-one matching charges for the change. The assignment excess is the extra cost of one-to-one matching over nearest neighbor on the same patches. Our identity is exact: the one-to-one score equals the nearest-neighbor score plus the excess. The identity localizes every ranking change between the two rules to a difference in excess between an anomalous and a normal image, so it shows which pairs the constraint corrects and which it reverses (Section 3.4). LogSAD and ObjectCore are related. They use Hungarian matching on regions or objects inside full systems, which addresses detection with their own components, not when the constraint itself helps (Appendix H). Our revised Introduction and Section 4.3 state the bank dependence, and we now define the patch multiset, assignment excess and coreset memory early.

w2: As a within-system control inside official LogSAD, assignment beats reuse in the branch, and without a segmenter its fused score is within -5.22 to +2.82 of LogSAD's region branch. As you asked, we compared under LogSAD's own settings. We kept its CLIP features, supports, calibration and fusion in the official four-shot system. On the same dense costs, our one-to-one assignment then beats reuse by +5.27 branch logical AUROC (95% range 1.73 to 8.81), positive in all five categories, on one support draw. After fusion the difference is +2.23, with an interval that includes zero. When we replace LogSAD's Segment Anything region branch with our mask-free dense assignment, fused logical AUROC is 85.80 against 87.01. Within that interval, from -5.22 to +2.82, the fused comparison ranks neither above the other (Appendix H, "Assignment inside LogSAD").

We also ran official LogSAD in its full-data setting (Appendix H, "Full-system comparison"). Our run reproduces its published combined AUROC, 90.13 against 90.2. LogSAD runs a reusable patch branch beside a Hungarian branch over Segment Anything regions. Our fusion adds one assignment signal to a trained host detector, EfficientAD-M, with no segmenter. At full data, our fusion minus LogSAD is -1.21 logical, from -8.40 to +4.41, and -1.50 structural, from -8.39 to +3.16, over five categories. So at full data, without a segmenter, we are too close to call against LogSAD. LogSAD and ObjectCore are system-level references whose encoders, data and components differ from ours. ObjectCore reports 80.8 combined MVTec LOCO AUROC in its four-shot setting, a different data regime from our full-data fusion. RAD (ICML 2026) reports no MVTec LOCO result, so there is no published number to compare with.

w3: At equal cardinality, forbidding reuse and covering the reference are one constraint, and the gain sits at that constraint. When the counts are equal, assigning every query patch without reuse necessarily covers every reference patch (Section 3.2). So our coverage controls compare objectives at equal cardinality, and our claim is the joint constraint, the one the abstract names. What our data show is where the gain sits. Allowing each reference patch a second use keeps 6% of the logical gain on DINOv2 and on DINOv3 (Table 8). From a cap of 12 the score equals nearest neighbor. So the gain is concentrated at the one capacity where every reference patch must be matched. Sections 3.2 and 3.3 and Appendix A.3 now state our claim this way, and our main contrasts (Tables 1 and 10a) need no split.

w4: The gain over Chamfer stays above zero on the hosts we trained, up to 82.51 logical AUROC, and our core result uses no host. You ask whether the branch mainly compensates for an undertrained host. Two of our results address it. Table 10a, our core result, uses frozen DINOv2 and DINOv3 features and no trained detector. And as the host improves, the gain persists. We trained three 70,000-update EfficientAD-M hosts with Imagenette penalty sampling. They reach 82.51 logical AUROC against the published 86.8, up from 76.56 for our 70-epoch host. On them our fusion reaches 88.08, against 85.06 with reuse fusion and 85.53 with Chamfer fusion. The gain over Chamfer is +2.55 across three seeds (95% range over seeds 1.38 to 3.72; Appendix H, "Longer host"). We compare fusion against reuse and Chamfer on the same host. Chamfer uses the same reference features, so the gain is what one-to-one matching adds beyond them.

Our host's shortfall from the published scores sits in two categories, pushpins and splicing_connectors. On the other three, where our host is within two points of the published score or above it, the gain over Chamfer is +3.52 (2.85 to 4.19) and over reuse +4.38. We chose those three by host score alone, after seeing the per-category results, so this is a post hoc check. So the gain over Chamfer is positive on both EfficientAD-M protocols we trained: +3.13 on the 70-epoch seeds and +2.55 on the longer-trained hosts. Appendix C.2 states it for those protocols.

I thank the authors for the additional experiments and analyses. The comparison between pooled and separate references makes the paper’s contribution clearer, but I still find it limited in scope. It remains unclear how well the benefit of one-to-one matching generalizes across backbones. For example, on VisA count anomalies, it improves AUROC by 1.87 points over nearest-neighbor matching with DINOv2, but reduces it by 0.84 points with DINOv3. In light of the new results, I have raised my score from 1 to 2.

Reviewer_pYRdOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the behavioural changes that arise when test patches are prohibited from reusing reference patches from normal samples. The authors conduct a controlled comparison between nearest-neighbor matching and one-to-one assignment, fixing the features, cost definitions and aggregation strategy across settings. One-to-one assignment improves performance on count- and composition-based logical anomalies, typically at the expense of structural anomaly detection. The authors systematically analyse this trade-off via assignment excess, reference construction designs, and integration with existing anomaly detectors.

Critical strengths
  1. The controlled comparison is valuable. By using an identical cost matrix for both matching schemes, the isolated effect introduced by the one-to-one matching constraint becomes straightforward to interpret. The ablation studies over coverage and aggregation further clarify the extent to which the performance gains on logical anomalies stem from the one-to-one requirement.
  2. The ablation on reference construction is especially insightful. Pooling patches from multiple normal images weakens the cardinality constraint, whereas evaluating each full reference image independently retains it. Experiments with incomplete references and contamination also deliver practical insights regarding best practices for reference bank design.
  3. The negative results effectively delineate the scope of the findings: no measurable improvement is observed for pure rearrangement anomalies; trends on count-based anomalies reverse on VisA under DINOv3; and integrating the assignment constraint into PatchCore leads to degraded structural performance.
Critical weaknesses
  1. A key limitation lies in the multi-seed comparison for the integrated detector. Table 4 presents gains over the base EfficientAD-M across five seeds, yet the comparisons against the reuse and Chamfer baselines are only reported for seed 0. While Appendix C.3 provides a supportive replication showing a 2.89-point gain over Chamfer even when retraining seed 0, paired comparisons across all five existing seeds would still be desirable. Such results would validate whether the assignment signal brings consistent, incremental benefits beyond simply incorporating an additional DINO reference feature.
  2. Appendix C.2 states an overly strong conclusion regarding the influence of host-model training. Under the 70-epoch training schedule, the host model achieves 76.56 / 87.69 logical / structural AUROC, whereas the official training schedule yields 86.8 / 94.7. Using this unified host ensures fair fusion comparisons, but it cannot demonstrate that the observed gains hold independently of the host’s training quality. I recommend limiting this claim strictly to the evaluated protocol.
  3. The reasoning is coherent for readers familiar with patch-based anomaly detection, yet dense domain jargon makes it less accessible to newcomers.
Questions for authors

Besides count anomalies, logical anomalies cover misplaced components and wrong ordering. It remains unclear whether the method behaves consistently across these logical anomaly subtypes. More discussion and results are expected.

3 comments

Across all five saved seeds, conditional excess beats both reuse and Chamfer. On all five saved 70-epoch seeds we ran the paired comparison you asked for. Conditional excess beats reuse by +3.64 logical points (95% range 3.42 to 3.85) and Chamfer by +3.13 (2.93 to 3.33). Both intervals, computed over training seeds on the fixed test set, lie above zero, and both gains are positive on every one of the five saved seeds (Appendix H, "Five saved seeds"). Averaged over seeds, structural AUROC stays within half a point of reuse and Chamfer. So our signal adds more than reuse or Chamfer does as a DINO reference feature.

You call the reference ablation "especially insightful". It shows that the same matcher helps or does not help depending on how the reference bank is built, and our revised main text now gives its numbers directly. On the jargon you mention, we now define the patch multiset and conditional excess in the Introduction, and coreset memory (a stored subset of normal patches) in Section 2. The gain comes from count and composition anomalies. In our exploratory, tag-based split, DINOv2 assignment gains +5.29 logical points over nearest neighbor on the logical anomalies outside the placement-only group (anomalies that only rearrange parts). On the 59 placement-only anomalies, in three categories, the difference is -0.38, with a paired 95% image-bootstrap interval from -1.72 to +0.92 that spans zero (Appendix H). A rearrangement that does not change the patch multiset leaves the score unchanged, so our claim now covers count and composition.

On a stronger, longer-trained host, fusion still leads Chamfer. Here we use three 70,000-update EfficientAD-M seeds trained with Imagenette penalty sampling, the trained host detector to which we add our signal. On these hosts, logical AUROC is 88.08, against 82.51 for the host alone. The gain over Chamfer is +2.55, with a 95% t-interval over the three paired seeds of 1.38 to 3.72, above zero (Appendix H, "Longer host"). So the gain over Chamfer is positive on both EfficientAD-M protocols we trained: +3.13 on the five 70-epoch seeds and +2.55 on these longer-trained hosts. The host's own score rises from 76.56 toward the published 86.8. Our claim compares fusion against reuse and Chamfer on the same host, and Appendix C.2 now states the gain for the protocols we evaluated, as you asked.

AuthorsOpenReview ↗

Follow-up (18 Nov): paired seeds, subtypes and the released checkpoint

Thank you again for your review. This follows up our 16 Nov reply with the paired numbers in one place and one new result. The table answers your two requests: paired comparisons across the saved seeds, and the logical subtypes. It also adds the released EfficientAD-M checkpoint, which speaks to your concern about host training.

Comparison (logical detection, points)Result
Five saved seeds, over reuse fusion (mean)+3.64, positive in every seed
Five saved seeds, over Chamfer fusion (mean)+3.13, positive in every seed
Released EfficientAD-M checkpoint, over Chamfer fusion+2.11 [1.01, 3.27]; structural 95.19 to 95.05
Placement-only anomalies, assignment minus nearest neighbour-0.38 [-1.72, 0.92]

So the gain is on count and composition, and it holds on the released checkpoint (Appendix H). We would welcome your assessment of these results.

Reviewer_pYRdOpenReview ↗

maintain my positive score.

Thank you for your responses. My main concerns are addressed, I maintain my positive score.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Overall response to reviews

Update: our latest summary is "Update (18 Nov): where each concern stands after the discussion".

Hungarian matching inside earlier full anomaly systems is known; what we add is that, on the same features and costs, the one-to-one constraint helps when each reference stays complete and gives almost nothing on a pooled bank. Reviewer GRp5 calls our paper "a clean controlled study of the correspondence constraint itself". Its main result is that the same one-to-one matcher helps or does almost nothing depending on how the references are stored. On DINOv2 it gains +5.69 logical AUROC points over nearest neighbor on separate complete references. On a pooled bank the difference is +0.24, too close to call (Table 10a).

We make one change to patch anomaly scoring and hold everything else fixed: each reference patch may be matched to only one test patch. We use Hungarian matching to enforce it. With the same five references and feature costs, merging them into one pooled bank leaves assignment almost no gain over nearest neighbor. Scoring each complete reference (one whole normal image kept intact) separately, with each rule keeping its minimum over the five, gives the large gain. Logical AUROC measures detection of anomalies in the number, composition or arrangement of parts, and structural AUROC measures detection of local defects. In our reply to 752b we give the single-reference result and the two propositions that explain why one-to-one matching charges for count changes that reuse misses.

This matters for anyone building a patch detector, because the patch detectors reviewed in Table 19, including the patch branch of an earlier full system, all use pooled patch banks. A one-to-one matcher tested only on a pooled bank can look ineffective even when it helps on complete references. Our advice for one-to-one patch scoring is to keep each reference complete and score it separately; Hungarian matching is the standard solver we use to measure this.

  • Full system (GRp5, 752b): our run reproduces the official system's full-data combined AUROC, and our patch assignment beats nearest neighbor on its own support images (one support draw).
  • Seeds (pYRd): across five saved seeds, conditional excess (the extra cost of one-to-one matching, minus that on similar normal images) beats both reuse and Chamfer (a two-way nearest-neighbor score).
  • Host training (pYRd, 752b): we add our signal to the score of a trained host detector, EfficientAD-M. On the longer-trained hosts (three seeds) our fusion beats both, and our Appendix C.2 now names the protocols we evaluated.
  • Scope (pYRd, GRp5, 752b): our claim now covers count and composition anomalies (extra or missing parts, or wrong kinds of parts), and the joint one-to-one constraint under complete references.
  • Routing and fusion (GRp5): category-agnostic retrieval chooses the same reference as correct-category retrieval on all 1,568 test images, and fusion raises logical AUROC on the EfficientAD-M hosts.

We collect the new results in Appendix H. Added to the EfficientAD-M detectors we trained, our signal improves logical detection in all five training seeds, and its gain over Chamfer holds on longer-trained hosts; the controlled result in Table 10a does not depend on host training.

AuthorsOpenReview ↗

Update (18 Nov): where each concern stands after the discussion

One-to-one patch matching helps when each normal reference is kept whole, and loses its gain when references are pooled.

Our main result is new empirical knowledge for practitioners: widely used detectors such as PatchCore keep a pooled bank of normal patches, and in our tests a pooled bank hides the gain, so one-to-one matching can look ineffective there.

This updates our 16 Nov overall response with the results added since, now in Appendix H of the revised PDF. Thank you to all three reviewers. Reviewer pYRd called the reference ablation "especially insightful", GRp5 called the paper "a clean controlled study of the correspondence constraint itself", and 752b noted the "practical significance" of the problem.

What is new (GRp5, 752b). In practice, score each complete normal reference separately. Adding our calibrated one-to-one signal, computed against complete normal references, to our AnomalyDINO reimplementation (exploratory, from cached features) improves logical detection; under DINOv3 the structural cost exceeds our one-point margin.

The gain holds on a released detector (pYRd, 752b). On the EfficientAD-M checkpoint released with PUAD, logical AUROC rises from 88.35 to 91.01, a gain of 2.11 over same-feature Chamfer fusion, while structural AUROC moves from 95.19 to 95.05 (Appendix H, "Released host").

The bank effect holds beyond one backbone (752b). It holds with WideResNet-50 and both DINOv2 and DINOv3, and on DigitAnatomy and VID-AD. In our separate-versus-pooled test on VisA with DINOv3, no gain is established.

Across seeds and subtypes, the gain is on count and composition (pYRd). Our reply reports the paired comparisons across saved seeds and the exploratory subtype split.

Appendix H of the revised PDF has these results. We would be glad to answer any further questions.

AuthorsOpenReview ↗

In brief (18 Nov): for anyone building on this work

A short version for readers who want only the takeaway; the details are in "Update (18 Nov): where each concern stands after the discussion" and in the reply threads.

TL;DR: match every patch once, against one whole normal image at a time.

Many anomaly detectors learn only from normal images. They split a test image into small patches and match each one to its closest patch in the normal images; a patch with no close match is flagged. Because many test patches may share the same normal patch, an extra, missing or wrong part can go unnoticed.

Requiring a one-to-one match, where each normal patch is used at most once, makes such changes in count and composition show up in the score. This works when each normal image is kept whole and scored on its own; pooling the normal images into one shared bank of patches nearly removes the gain.

Widely used detectors such as PatchCore keep a pooled bank of this kind, so one-to-one matching can look ineffective there. Across several encoders and datasets, keeping each normal image whole preserves this gain, and in our tests adding the signal to an existing detector improves its detection of these count and composition anomalies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.