acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Normal Memory Readout for Training-Free Anomaly Detection

Abstract

Normal patch memories support anomaly detection by comparing test features with stored normal references. Nearest-neighbor scoring measures proximity to one reference, whereas nearby references can jointly provide a local estimate of normal appearance. We propose ProCon, a training-free anomaly detector that retrieves normal neighbors, reconstructs a query by their distance-weighted average, and uses the reconstruction residual as its anomaly score. Independent memories at different DINOv2 layers preserve the normal neighborhood of each feature space; a median across memory banks followed by a mean across layers combines their residual maps. With the encoder and memory fixed, replacing nearest-neighbor distance with local reconstruction improves pixel AP by 1.4 and 2.5 percentage points on MVTec-AD and VisA. Under the Dinomaly evaluation protocol, ProCon achieves pixel AP of 75.3% and 57.0%, respectively, outperforming the compared Dinomaly and INP-Former configurations on this metric. At matched memory storage, independent layer reconstruction improves pixel AP and AUPRO over feature concatenation in all 27 categories. These results show that a normal patch memory can serve as an effective local reconstruction model without training a decoder.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
anomaly detection, anomaly localization, training-free learning, memory banks, local reconstruction, pretrained vision models

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_V9SJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ProCon, a training-free method for industrial anomaly detection using frozen DINOv2 features. Instead of scoring a query by its distance to the nearest normal reference, it reconstructs the query from several distance-weighted normal neighbors and uses the reconstruction residual as the anomaly score. Separate memories are maintained at different encoder layers, with residuals aggregated across memory banks and layers. Experiments on MVTec-AD and VisA include comparisons with Dinomaly and INP-Former, same-memory scoring ablations, matched-storage comparisons, and inference-cost measurements.

Critical strengths

The same-memory comparison provides clear evidence for the proposed readout. Holding the encoder and stored references fixed, replacing nearest-neighbor distance with reconstruction residual improves pixel AP by 1.4 and 2.5 percentage points on MVTec-AD and VisA, respectively (Table 3). This isolates the benefit of scoring from changes in memory content.

The matched-storage comparison supports independent retrieval across layers. In the reported seed-0 experiment, independent reconstruction improves pixel AP and AUPRO over feature concatenation in all 27 categories (Table 4 and Appendix F). This is useful evidence that memory organization matters beyond its storage budget.

Critical weaknesses

The methodological novelty appears limited. The main design combines local neighbor reconstruction with layer-wise score aggregation. Tables 10–11 show that uniform and distance-weighted five-neighbor reconstruction differ by less than 0.1 percentage point in pixel AP on both datasets. Thus, the distinctive weighting rule contributes little to the reported gain, and the advance over existing reconstruction-based approaches needs a clearer justification.

The external comparisons do not adequately establish the contribution against the closest alternatives. PatchCore and FastRecon are discussed but not directly evaluated under comparable settings. In addition, the layers and aggregation rule were selected using MVTec-AD results without clearly specifying a separate selection split. These gaps limit confidence in the comparative advantage and in the MVTec-AD performance estimate.

Questions for authors
  1. Can the authors provide comparisons with PatchCore and FastRecon using comparable backbone features, normal training data, and storage budgets where applicable? These comparisons would help distinguish the benefit of the proposed readout from improvements due to the feature extractor and implementation choices.

  2. Which data and labels were used to select the layers and aggregation rule on MVTec-AD? Was selection independent of the reported evaluation set? If not, can the authors provide an evaluation with a separate selection split or held-out categories?

2 comments
AuthorsOpenReview ↗

Response to Reviewer V9SJ

We thank the reviewer for recognizing the same-memory and matched-storage comparisons. We clarify the contribution beyond distance weighting, add the requested external baselines, and make configuration selection explicit.

Q1. Contribution and comparisons with PatchCore and FastRecon

Our question is how normal-memory readout and layer-wise retrieval affect localization with frozen features. The revision separates these choices rather than attributing the complete detector's improvement to its weighting rule.

With identical memory, mean k-NN improves P-AP over 1-NN by 1.15 points on MVTec-AD and 1.92 on VisA after four-layer aggregation. Both centroid readouts achieve higher mean P-AP than mean k-NN, while mean k-NN has the highest AUPRO. Uniform and weighted centroids have similar aggregated P-AP; weighting helps most at layer −3 (Tables 3, 11, and 13).

Memory organization addresses a separate question. At exactly matched stored feature bytes in every category, independent reference selection and retrieval at each layer outperform selection and retrieval in a concatenated feature space in P-AP in all 27 categories for each of five seeds (Table 5, Figure 2, Table 14).

The contribution is controlled evidence about which retrieval and scoring choices matter and when they help. We compare the resulting detector with the requested baselines:

MethodFeature/data conditionsMVTec P-AP/AUPROVisA P-AP/AUPRO
DINOv2-PatchCoreSame four blocks; 1024-D FP32; matched stored bytes66.58/91.6044.78/82.64
FastRecon, public-code configurationWRN50, 8-shot55.14/82.0630.95/71.18
FastRecon, DINOv2 adaptationFour pooled/concatenated layers; all normal training images as candidates65.72/88.3426.47/69.02
ProCon, B=1Independent four-layer memories; 512-D FP1675.00/95.0656.59/93.78

DINOv2-PatchCore and ProCon use the same four encoder blocks and exactly matched stored feature bytes per category: 111.1 MiB across MVTec-AD and 265.2 MiB across VisA. ProCon's category-paired P-AP gains are 8.41 and 11.81 points, with 95% CIs of [+4.03,+13.52] and [+4.55,+19.17]. Feature aggregation and storage precision differ, so this evaluates the complete detectors. Readout effects are assessed separately with identical memory tensors.

We evaluated FastRecon's public WRN50 configurations at 1/2/4/8 shots and a DINOv2 adaptation. The adaptation uses the same four DINOv2 layers, applies 3×3 average pooling, concatenates them into 2048 dimensions, and uses all normal training images as candidates. With public parser defaults ρ=0.001\rho=0.001 and λ=2.0\lambda=2.0, it stores approximately one-tenth as many feature values as ProCon B=1. Appendix F.1 states this storage difference and the differences between the public-code and original-paper settings.

Q2. Configuration selection and presentation

The default layers and aggregation rule were selected using MVTec-AD test images and their image labels and pixel masks, without a separate validation split. The configuration was applied unchanged to VisA and Real-IAD.

We assessed sensitivity through retrospective reselection among 170 candidates under image half-split, category-held-out, and cross-dataset protocols. Each selection split excludes the images, category, or dataset used for its corresponding evaluation. MVTec-AD P-AP ranges from 73.91% to 74.84%. Selection on MVTec-AD before VisA evaluation recovers the default for all three seeds (Table 8).

These retrospective comparisons measure selection sensitivity; they do not make the original MVTec-AD evaluation independent of the initial configuration choice.

Real-IAD was excluded from development and configuration selection. The fixed B=5 configuration achieves 51.72±0.02% P-AP and 95.05±0.02% AUPRO across three seed groups, covering all 30 categories and five camera views. Aggregate and category-wise results appear in Tables 1 and 9.

We reorganized the introduction and experiments around readout and memory organization and consolidated external comparisons in Table 2. Both B=1 and B=5 combine four layers; B=5 additionally aggregates five banks within each layer.

I've seen the author's response regarding the contribution writing and the additional baseline experiments. First of all, thank you very much for providing those extra baseline comparisons.

However, I believe that if this manuscript is to be accepted, the writing still needs considerably more polish. The contributions and the methodology need to be explained much more clearly. Because of this, I am maintaining my original score.

I think the authors should take the time to thoroughly refine the writing and experiment before submitting it elsewhere. Therefore, I suggest that they use the reviewers' feedback from this round to put together a stronger version and make the logical flow of the writing much tighter.

Reviewer_JNV7OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ProCon, a training-free method for anomaly detection based on a normal feature memory. Instead of measuring the distance to the nearest normal feature, it reconstructs a query feature from several normal neighbors and uses the reconstruction residual as the anomaly score. The method uses separate memories at four DINOv2 layers, takes the median over memory banks and the mean over layers. It is evaluated on MVTec-AD and VisA. The paper compares Hard and Soft readout with the same memory, studies memory organization with matched storage, and reports inference cost. The readout change improves pixel AP by 1.4 and 2.5 percentage points on the two datasets.

Critical strengths
  1. The Hard/Soft comparison. I appreciate the comparison with the same memory. The encoder, projection, reference features, and number of layers are fixed, so it is easier to see the contribution of the reconstruction residual. Soft readout improves all four pixel metrics on both datasets. The paper also separates this improvement from the larger gap to Dinomaly and INP-Former, which makes the results easier to understand.

  2. The analysis of memory organization. The independent and concatenated variants are compared with the same feature storage. Independent reconstruction improves pixel AP and AUPRO in all 27 categories. The one-bank and five-bank results also show how much the ensemble helps and how much it costs. The supplement provides the core code and explains that it is not a complete benchmark reproduction package.

Critical weaknesses
  1. The model selection protocol is unclear to me. Sec. 4.1 and Appendix C.1 mention that the layers and aggregation rule are selected using MVTec-AD results, considering both detection and localization. I did not find whether a separate validation split is used, or whether the selection uses the test labels in the reported benchmark. This detail is important for interpreting the MVTec-AD results, especially since the method is training-free. Applying the same configuration to VisA is useful, but it does not answer how MVTec-AD is used during selection. I am not concluding that test leakage happens here, but would like to have this detail clarified.

  2. The comparison with previous reconstruction methods. The paper discusses FastRecon, but the main comparisons are with Dinomaly, INP-Former, and the internal nearest-neighbor variants. Thus it is a little bit hard to understand the additional contribution over existing training-free reconstruction. The ablations also show that uniform reconstruction with five neighbors has almost the same pixel AP as the proposed weighting, and five banks only give a small improvement over one bank. To me, the useful part seems to be the local reconstruction residual with independent layer memories. A direct comparison under the same feature and data setting would help show this contribution more clearly.

Questions for authors
  1. Which images, category splits, and labels are used to select the layers and aggregation rule? Are any of the reported MVTec-AD test labels used? If so, please separate the selection results from an untouched evaluation and provide a held-out selection protocol.

  2. Would it be possible to compare with a relevant training-free reconstruction method, such as FastRecon, using the same encoder, normal data, and evaluation? I understand that FastRecon is originally for few-shot use, so directly comparing the published numbers with different data access would not resolve this question.

  3. Appendix G.5 mentions that normal neighbors could also reconstruct an anomalous query and make its residual small. How often does this happen in the benchmark errors? It would be helpful to compare the false negatives of Soft and Hard to understand when the proposed readout helps or hurts.

1 comment
AuthorsOpenReview ↗

Response to Reviewer JNV7

We thank the reviewer for recognizing the same-memory and matched-storage comparisons. We clarify configuration selection, add FastRecon comparisons, and quantify when centroid reconstruction helps or hurts detection.

Q1. Configuration selection and subsequent evaluation

The default layers {−3,−6,−8,−9}\{-3,-6,-8,-9\}, unweighted layer averaging, and aggregation without map normalization were selected using MVTec-AD test images and their image labels and pixel masks, without a separate validation split. This configuration was applied unchanged to VisA and Real-IAD.

We retrospectively reselected among 170 candidates spanning layer subsets, aggregation, and normalization. Each selection split excludes the images, category, or dataset used for the corresponding evaluation. Appendix C specifies candidates and splits.

EvaluationSelection/evaluation protocolP-AP
MVTec-ADSelect on one image half; evaluate on the other74.84
MVTec-ADFixed default on the same evaluation halves75.07
MVTec-ADSelect on 14 categories; evaluate on the held-out category73.91
MVTec-ADSelect on VisA74.40
VisASelect on MVTec-AD56.59
VisASelect on 11 categories; evaluate on the held-out category57.88

Results use B=1 and seeds 0/1/2 (Table 8). The default is recovered in 4/6 image half-splits and 31/45 category-held-out folds on MVTec-AD, and for all three seeds when selecting on MVTec-AD before evaluating VisA. These analyses measure selection sensitivity retrospectively and are distinguished from the initial use of MVTec-AD test performance.

Real-IAD was excluded from development and configuration selection. The fixed B=5 configuration covers all 30 categories and five views. Across non-overlapping bank-seed groups 0–4, 5–9, and 10–14, it achieves 91.80±0.02% I-AUROC, 51.72±0.02% P-AP, and 95.05±0.02% AUPRO.

Q2. Bank aggregation and FastRecon

Both configurations use four independent layer memories. B=1 uses one bank per layer; B=5 aggregates five banks per layer by pixel-wise median.

DatasetB=1 P-APB=5 P-APB=1 latencyB=5 latency
MVTec-AD75.0075.2620.9 ms39.8 ms
VisA56.5956.9922.2 ms45.4 ms

Latency is per image. One bank per layer already approaches five-bank P-AP. Additional banks give modest gains and reduce seed variability on VisA at increased cost (Tables 1 and 6). Table 3 also reports the similar aggregated P-AP of uniform and weighted centroids.

We evaluated FastRecon's public WRN50 few-shot settings and a DINOv2 adaptation using the same four layers and all normal training images as candidates. The adaptation applies 3×3 average pooling per layer and concatenates features into 2048 dimensions. With public parser defaults ρ=0.001\rho=0.001, λ=2.0\lambda=2.0, it achieves 65.72/26.47% P-AP on MVTec-AD/VisA, versus ProCon B=1's 75.00/56.59%.

This FastRecon configuration stores approximately one-tenth as many feature values as ProCon B=1. Table 2 and Appendix F.1 state the processing and storage conditions. Equal-byte memory organization and identical-reference readout effects are assessed by the separate controlled comparisons.

Q3. Small anomalous residuals and false negatives

An anomalous query near its normal neighbors' centroid can receive a small residual. We compared Hard (1-NN) and Soft (weighted centroid) at matched normal-pixel FPR, with thresholds calibrated separately per readout and category.

At FPR = 0.05, the fractions of defect pixels missed exclusively by one readout are (B=1, seeds 0/1/2):

DatasetHard misses, Soft detectsSoft misses, Hard detects
MVTec-AD2.77%0.55%
VisA3.76%0.30%

Soft additionally detects more defect pixels than it newly misses.

A connected region is detected when at least half its pixels exceed the threshold (Table 4). For seeds 0/1/2, Soft-only detections exceed Hard-only detections at all five FPRs on both datasets. Some MVTec-AD differences are small and reverse with seeds 0/5/10: at FPR = 0.05, Hard-only and Soft-only detections are 0.59% and 0.55%.

For regions smaller than 64 pixels at FPR = 0.05, mean k-NN exceeds ProCon in PRO: 86.54 versus 84.73% on MVTec-AD and 87.67 versus 86.08% on VisA (Appendix G.2).

Separately, we measured q(z)=∥z−z^∥22∑jwj∥z−mj∥22.q(z)=\frac{\|z-\hat z\|_2^2}{\sum_j w_j\|z-m_j\|_2^2}. At most 0.14% of defect patches fall below the normal-patch fifth percentile of this ratio across evaluated layers and datasets. This is the frequency of a specified low-ratio event; actual missed detections are assessed by the matched-FPR comparisons above.

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

This paper focuses on industrial anomaly detection and investigates the question of using the normal patch memory as a local reconstruction model for anomaly scoring. The paper proposes ProCon, a training-free anomaly detector that retrieves normal neighbors, reconstructs a query by their distance-weighted average, and uses the reconstruction residual as its anomaly score. Extensive experiments evaluate the effectiveness of the proposed ProCon for industrial anomaly detection comprehensively.

Critical strengths
  1. Comprehensive and well-controlled ablation studies: Each designed module is well ablated, providing a comprehensive effectiveness analysis for anomaly detection.
  2. Simple, efficient, and easy to reproduce: The method needs no training and no decoder, and all layers and banks share one encoder forward pass.
Critical weaknesses
  1. Limited novelty of the distance-weighted readout: The Soft readout (Eq. 3-4) seems to be Nadaraya-Watson kernel regression in feature space, essentially, and the paper does not discuss this connection. Moreover, the paper's own ablation shows that the distance weighting contributes little in Tables 10 and 11. The gain therefore seems to come mainly from using k>1 neighbors rather than from the weighting. Since k-NN averaging is already used in prior work (e.g., SPADE [1]), the remaining difference is scoring against the neighbors' weighted center instead of averaging the distances to them, which corresponds to the dispersion term in Eq. 12. This difference is not isolated by any experiment.
  2. AUPRO performance: In Table 1, ProCon's AUPRO is lower than the best baseline on both datasets (95.2 vs. 95.4 for INP-Former on MVTec-AD; 94.0 vs. 95.1 for Dinomaly on VisA). The main text reports these numbers (L308) but does not acknowledge or explain the gap, so the advantage over existing methods rests on pixel AP alone.
  3. Missing baselines: AnomalyDINO [2] uses the same frozen DINOv2 patch features with nearest-neighbor scoring and is training-free, which makes it the closest prior work, but it is neither cited nor compared. Training-free methods such as MuSc [3] should also be included as a reference.

[1] Cohen, Niv, and Yedid Hoshen. "Sub-image anomaly detection with deep pyramid correspondences." arXiv preprint arXiv:2005.02357 (2020). [2] Damm, Simon, et al. "AnomalyDINO: Boosting Patch-based Few-Shot Anomaly Detection with DINOv2." WACV 2025. [3] Li, Xurui, et al. "MuSc: Zero-shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images." ICLR 2024.

Questions for authors
  1. How does the Soft readout relate to kernel regression? Could the authors add a k-NN average-distance baseline (mean of ∣∣z−mj∣∣||z - m_j|| over the same k=5 neighbors) to isolate the effect of reconstructing the weighted center?
  2. Why is ProCon's AUPRO lower than the best baseline while its pixel AP is higher? It would be helpful to provide a clearer analysis of the result.
  3. Please consider comparing with AnomalyDINO and other training-free methods such as MuSc.

Minor question: The full name of "ProCon" is never given; please spell it out at first use.

2 comments
AuthorsOpenReview ↗

Response to Reviewer tjvi

We thank the reviewer for recognizing the controlled ablations and the method's simplicity and efficiency. We address the requested comparisons and analyses below.

Q1. Kernel regression, mean k-NN, and distance weighting

As you noted, Eqs. (3)–(4) correspond to local Nadaraya–Watson regression over retrieved neighbors with identity targets yj=mjy_j=m_j. We now explain this connection in Related Work and Section 3.2 and cite Nadaraya (1964) and Watson (1964).

We fixed the encoder, projection, and memory tensors and compared four readouts. All multi-neighbor readouts use the same five neighbors. Results combine four layers with one bank per layer, averaged over seeds 0, 1, and 2 (Table 3).

ReadoutMVTec P-AP/AUPROVisA P-AP/AUPRO
1-NN73.60/94.0554.07/89.69
Mean k-NN74.75/95.0755.99/94.31
Uniform centroid74.96/95.0656.61/94.14
Weighted centroid75.00/95.0656.59/93.78

Multiple neighbors account for a substantial part of the gain over 1-NN. Both centroids have higher mean P-AP than mean k-NN, whereas mean k-NN has the highest AUPRO.

To isolate reference dispersion, we also compared the centroid residual with RMS neighbor distance using identical neighbors and weights:

\sqrt{\sum_j w_j\|z-m_j\|_2^2}.$$ The centroid residual has higher mean P-AP in all four comparisons (two weighting rules × two datasets), but all category-paired 95% confidence intervals include zero (Table 11). We distinguish these modest differences from the larger improvement over 1-NN. Weighting has a clearer effect at the deepest selected layer, −3: it improves P-AP over the uniform centroid by 1.03 points on MVTec-AD and 2.54 on VisA. At this layer on VisA, only the weighted centroid among the multi-neighbor readouts exceeds 1-NN. Tables 3 and 13 report all selected layers. **Q2. Higher pixel AP but lower AUPRO** We analyzed PRO at matched normal-pixel FPR and by defect size under the common GT>0 protocol (Tables 20–21). Below, differences are ProCon ($B=5$) minus the comparator, in percentage points. | Dataset/comparator | ΔP-AP | ΔAUPRO | Δ[email protected] | Δ[email protected] | Δ[email protected] | |---|---:|---:|---:|---:|---:| | MVTec/Dinomaly | +6.42 | +0.34 | +8.22 | +0.14 | −0.80 | | MVTec/INP-Former | +4.75 | −0.44 | +6.47 | −0.54 | −0.92 | | VisA/Dinomaly | +6.15 | −1.12 | +5.98 | −2.17 | −0.61 | | VisA/INP-Former | +8.25 | +0.64 | +5.01 | +1.54 | −0.00 | At FPR = 0.001, all four category-paired 95% confidence intervals for ΔPRO are positive. More permissive FPRs favor learned models in some comparisons. On VisA at FPR = 0.05, ProCon's PRO difference from Dinomaly is −4.05 points for regions smaller than 64 pixels, versus −0.19 and −0.46 for medium and large regions. Tiny defects constitute approximately 46% of VisA's connected defect regions. Pixel AP evaluates pixel-level precision and recall; AUPRO weights connected regions equally and integrates overlap across an FPR range. Better pixel ranking therefore need not imply better region coverage. The analyses locate ProCon's advantage at strict false-positive constraints and identify its limitations on some operating points and tiny defects. **Q3. AnomalyDINO and MuSc** We added both baselines to the common evaluation pipeline (Table 2; Tables 17–18). | Method/configuration | MVTec P-AP/AUPRO | VisA P-AP/AUPRO | |---|---:|---:| | AnomalyDINO, ViT-B/14, 392, full-shot | 61.09/92.22 | 51.56/91.68 | | MuSc, DINOv2 ViT-B/14, 518, zero-shot | 70.64/94.42 | 44.26/90.05 | | ProCon, ViT-B/14, 392, B=1 | 75.00/95.06 | 56.59/93.78 | Full-shot AnomalyDINO stores all normal training patch features; MuSc compares unlabeled test images. We state their resolution, data-access, and memory differences. Additional AnomalyDINO backbone/few-shot and MuSc backbone results appear in the appendix. ProCon has higher mean P-AP on both datasets. On MVTec-AD, its category-paired P-AP difference from MuSc has a 95% CI of [−1.36,+10.84] points, including zero (Appendix F.2). **Q4. Full name** We spell out ProCon as Projection-Consistency Memory in the abstract, introduction, and Algorithm 1.

I thank the authors for the detailed response and the revised paper. The added comparisons with AnomalyDINO and MuSc, and the discussion of the AUPRO gap, address W2 and W3.

For W1, the new same-memory controls are helpful, but they confirm my concern. Mean k-NN already obtains most of the P-AP gain over 1-NN, distance weighting adds little in the four-layer setting, and the centroid-versus-RMS intervals all include zero. As the revised paper itself states, the evidence supports an advantage over 1-NN more clearly than over other multi-neighbor readouts. The remaining contribution is mainly the use of independent per-layer memories, which I find limited for ICLR.

Therefore, I maintain my original score.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Clarifying the Contribution and Main Findings

We thank all reviewers for their careful assessment and constructive feedback, particularly their recognition of the controlled comparisons. We clarify below how the empirical findings support the contribution beyond the novelty of individual components.

1. What explains the readout improvement?

We thank the reviewer for identifying the connection to Nadaraya–Watson regression, which we had not explicitly recognized in the original manuscript. The distance-weighted centroid in Eqs. (3)–(4) can be interpreted as local kernel regression with identity targets yj=mjy_j=m_j. We now state this connection and cite the original work.

With the encoder, projection, and memory fixed, 1-NN, mean k-NN distance, uniform centroid, and distance-weighted centroid achieve P-AP of 73.60/74.75/74.96/75.00% on MVTec-AD and 54.07/55.99/56.61/56.59% on VisA. Thus, multiple neighbors account for a substantial part of the improvement; centroid reconstruction and weighting provide smaller, condition-dependent differences.

Weighting matters more at layer −3: it improves P-AP over the uniform centroid by 1.03 and 2.54 percentage points on MVTec-AD and VisA. At this layer on VisA, mean k-NN and the uniform centroid fall below 1-NN, whereas the weighted centroid exceeds it. The small difference after layer aggregation therefore does not describe every layer.

2. Research question and contribution

Our guiding question is how much localization performance can be obtained from frozen representations by changing how normal references are organized and read, without learning a reconstruction decoder. Understanding these choices requires separating their effects from the performance of the complete detector.

We therefore examine two design choices through distinct controls: readout with identical memory tensors, and memory organization with matched stored feature bytes. The contribution is the resulting evidence about which choices matter, when they help, and where their benefits stop.

At matched stored bytes in every category, independent layer memories outperform concatenated memory in P-AP in all 27 categories for each of five seeds, with dataset-average gains of approximately 1.8–2.5 points. Independent four-layer memory also exceeds a single-layer 4% coreset: 74.99 versus 73.97% P-AP on MVTec-AD and 56.44 versus 50.39% on VisA. These results support preserving each layer's normal neighborhood and combining independently retrieved evidence, rather than concentrating references in one representation or searching a joint feature space.

Independent memories and late fusion are established concepts. Our contribution is to quantify their consequences for normal-memory anomaly detection under controlled conditions, alongside the smaller and less consistent effects of alternative readouts.

3. What do the localization metrics show?

ProCon achieves higher mean pixel AP than the evaluated training-free and learned baselines, but not the highest AUPRO in every comparison. Matched-FPR and defect-size analyses identify where the rankings differ.

At normal-pixel FPR = 0.001, ProCon with B=5B=5 exceeds Dinomaly and INP-Former in PRO by 8.22 and 6.47 points on MVTec-AD, and 5.98 and 5.01 points on VisA. At more permissive FPRs or on tiny defects, some alternatives obtain higher region overlap. These findings characterize the operating conditions under which local reconstruction is effective.

4. Scientific and practical significance

With one bank per layer (B=1B=1), ProCon achieves 75.00% P-AP on MVTec-AD and 56.59% on VisA. At matched stored feature bytes, the category-paired gains over DINOv2-PatchCore are 8.41 and 11.81 points. Its mean P-AP also exceeds the evaluated full-shot AnomalyDINO configurations, Dinomaly, and INP-Former.

Applying the configuration selected on MVTec-AD unchanged to all 30 categories and five views of Real-IAD gives 91.80% I-AUROC, 51.72% P-AP, and 95.05% AUPRO with B=5B=5.

The central finding is that normal feature memory can serve as an effective local reconstruction model without a learned decoder, and that its organization and readout materially affect localization. The controlled analyses explain the contributions and limits of these choices; the competitive performance demonstrates their practical relevance. We do not attribute the entire improvement to distance weighting or claim superiority under every metric.

AuthorsOpenReview ↗

Notification of Paper Title Change

We have updated the paper title to "ProCon: Local Reconstruction from Normal Memory for Visual Anomaly Detection" to better reflect the central contribution of the revised manuscript. The abstract and manuscript have been updated accordingly.

Dear Reviewers,

The authors have posted responses. As the discussion period is ending soon, could you please take a look and reply shortly, or at least acknowledge that you have read it?

Thank you!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.