acceptodds
Under review as a conference paper at ICLR 2027

GRef: Graph-Guided Reference Selection for Generalizable Industrial Anomaly Detection

Abstract

Frequent product changes make repeated training of industrial anomaly detectors costly, motivating methods that generalize to unseen products using normal reference images. When the available normal pool exceeds the reference budget, detection performance depends on which images are selected. Retrieving the most similar images individually can produce redundant references that provide limited complementary evidence. We propose GRef, a graph-guided framework that combines transferable relational representations with reference selection balancing query relevance and redundancy. It supports both selector-guided detector training and plug-in inference without target-side parameter updates. We establish a generalization bound that accounts for discrete reference selection, relating detection error to selection margins and model complexity. Across six industrial datasets, ResAD++ trained with our selected references achieves an average Image AUROC of 87.94%, outperforming all compared learned detectors. Plug-in integration with UniVAD achieves the best overall average of 93.70%. With detector parameters fixed, GRef also achieves the highest average performance among the compared selection strategies for all four evaluated detectors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
industrial anomaly detection, generalizable anomaly detection, reference selection, graph-based representation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Lqe3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

GRef combines a regional graph encoder with relevance–diversity reranking to select query-specific normal references for generalizable anomaly detection. However, I think the biggest problem with this paper is the setting and motivation: setting the budget based on the number of images makes no sense, and the observation in Fig. 1 does not justify the use of a graph. The experimental gains are very small, and there is no variance, so this method is not well supported.

Critical strengths

The comparisons across four fixed detectors provide useful controlled evidence that GRef improves reference selection under the prescribed reference budget.

Critical weaknesses
  1. Insufficient justification of the setting. Why must exactly K reference images be selected from a pool of normal samples? Limited compute does not by itself imply that the budget should be a fixed number of images, so this setting seems to have been invented to accommodate the authors’ own retrieval pipeline.
  2. Gap between the observation and the proposed solution. Fig. 1 only shows that the quality of the selected reference images affects the results, but it does not prove that the bottleneck is redundancy or spatial structure. Existing methods already perform feature matching against the query, yet the authors do not clearly state what problem remains or why a graph can solve it.
  3. Insufficient validation. The encoder learns dropout consistency and similarity preservation, while MMR penalizes embedding similarity. Neither has been shown to measure how useful a reference image is for detection or whether it can provide complementary evidence.
  4. The gains are very small. With UniVAD, the strongest detector tested, GRef is only 0.09 points higher than CPR. After training, ResAD++ improves by only 0.17 points, while RegAD drops by 0.34 points. No variance across multiple runs is reported, so it is unclear whether these differences are reliable.
Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Lqe3 (Part I)

Dear Reviewer Lqe3,

Thank you for your feedback on the four fixed detectors and for your patience with our delayed reply. We took the extra time to test a simpler selector with more references and to benchmark GRef against the full feature bank and its coreset.

Response to Weakness 1 (reference budget in deployment)

Section 1 gives the example of a ceramic bottle line changing designs. As production ramps up, the line can quickly accumulate many images of normal bottles. A reference-based detector can use these images to inspect the new design without retraining at each changeover.

Even with cached features, matching each bottle against regional features from the entire pool requires more computation and data access as the pool grows. Many normal images also show similar patterns on the bottle or similar viewing angles. For each query, GRef selects a few references that are relevant in appearance and have little redundancy, then passes them to the detector for fine-grained matching.

K is the number of references passed to the detector for each query. It can vary with the product and the latency budget. We study the regime where K is much smaller than the normal pool. Fixing K in the experiments controls the detector’s per-query reference workload and gives all selectors the same budget. Prior few-shot evaluations of RegAD, ResAD++, InCTRL, SPADE, PaDiM, PatchCore, and D²4FAD have reused a randomly sampled set of KK normal references across queries, which motivates us to ask whether choosing the KK references for each query can improve detection.

Response to Weakness 2 (motivation for graph encoding and MMR)

In Figure 1, the same detector uses four references from a pool of eight normal images. Changing only the selection strategy changes image-level AUROC on the test set. Existing query-adaptive matching methods can find features in the supplied references that correspond to regions in the query. When the normal pool is much larger than the reference budget, the selection of images determines which normal patterns are available for this matching step.

Graph encoding estimates the relevance of each reference to the query. Product appearance and texture can change from one product to the next, while spatial adjacency and cross-scale containment have the same definition across products. GRef encodes regional appearance together with these relationships, allowing retrieval to distinguish candidates with similar local textures but different spatial arrangements. Related approaches to learning transferable representations include the multiscale relational modeling in UniOD [1] and the encoding of structure and context in REFI-GAD [2].

MMR reduces redundancy in the selected set. Several nearly identical images can rank highly by relevance. MMR penalizes candidates that resemble an already selected reference, leaving room for other normal patterns. With the graph encoder fixed, MMR raised mean image-level AUROC from 87.52% to 87.94%, a gain of 0.42 percentage points over relevance-only Top-K. With the same eight reference slots, accounting for overlap among references gave the detector a better set than ranking by relevance alone.

[1] Dazhi Fu and Jicong Fan. UniOD: A Universal Model for Outlier Detection across Diverse Domains. ICLR 2026.

[2] Yujing Liu et al. Rethinking Feature Alignment in Generalist Graph Anomaly Detection: A Relational Fingerprint-based Approach. ICML 2026.

AuthorsOpenReview ↗

Response to Reviewer Lqe3 (Part II)

Response to Weakness 3 (reference utility and redundancy proxies)

Appendix E.1 reports the comparison of Top-K and MMR in Table 4 and the ablations of the two training losses in Table 5.

A reference can match the query well and still repeat a normal pattern covered by a selected reference. With the graph encoder fixed and eight references per query, penalizing similarity to already selected references improved all three detection and localization metrics.

Q1: How do the two training objectives support reference selection?

GRef encodes regional appearance, spatial adjacency and containment across scales. The similarity-preservation loss encourages the embeddings to retain similarities in the original regional features. Dropout consistency encourages the same image to have similar embeddings under different feature masks. We removed each loss in turn and retrained the encoder with the reference pool, budget, MMR rule and ResAD++ weights fixed.

Training objectiveImage-level AUROC (%)Pixel-level AUROC (%)AUPRO (%)
Without dropout consistency87.55 ± 0.0592.95 ± 0.0184.90 ± 0.05
Without similarity preservation87.40 ± 0.0592.90 ± 0.0284.75 ± 0.07
Both loss terms87.94 ± 0.0093.08 ± 0.0185.34 ± 0.03

We averaged the six datasets equally within each seed and report the mean ± sample standard deviation over three seeds. Values shown as ±0.00 are rounded.

Removing dropout consistency or similarity preservation lowered mean image-level AUROC by 0.39 and 0.54 percentage points, respectively. References selected with both losses gave the highest scores on all three metrics.

Q2: Why use embedding similarity as a proxy for reference redundancy?

A close match to even one selected reference suggests that the candidate may repeat a regional pattern already covered. We kept the encoder, candidates, detector weights and reference count fixed. Top-K (λ = 1) ranks candidates by query relevance. MMR (λ = 0.7) also penalizes each candidate’s highest embedding similarity to a selected reference.

Metric (%)Relevance-only Top-K (λ = 1)MMR (λ = 0.7)Mean paired gain (percentage points)
Image-level AUROC87.52 ± 0.0487.94 ± 0.00+0.420
Pixel-level AUROC92.91 ± 0.0293.08 ± 0.01+0.173
AUPRO85.00 ± 0.0585.34 ± 0.03+0.338

Averaged over three paired seeds, MMR gained 0.420, 0.173 and 0.338 percentage points over Top-K in image-level AUROC, pixel-level AUROC and AUPRO, respectively. At the same eight-reference budget, similarity to the selected set added useful information beyond query relevance and led to better detection and localization.

AuthorsOpenReview ↗

Response to Reviewer Lqe3 (Part III)

Additional results on graph encoding and reference selection

Table 4 compares full GRef with raw regional features plus MMR, a graph-free MLP plus MMR, and the graph encoder with Top-K.

Full GRef reached 87.94% mean image-level AUROC across six datasets. It gained 1.74, 1.67 and 0.42 percentage points over raw regional features with MMR, a graph-free MLP with MMR, and the graph encoder with Top-K, respectively. Keeping both graph message passing and MMR gave better detection than each of the three alternatives we tested. Within each seed, all configurations shared the same ResAD++ weights, normal pool and test set, and selected 8 references from 32 candidates.

Selection ruleRepresentationImage-level AUROC (%)Pixel-level AUROC (%)AUPRO (%)
Top-KGlobal pooling85.55 ± 0.0591.80 ± 0.0283.82 ± 0.06
Raw regional features85.95 ± 0.0591.90 ± 0.0284.03 ± 0.06
Graph-free MLP86.02 ± 0.0591.99 ± 0.0284.17 ± 0.05
Graph encoder87.52 ± 0.0492.91 ± 0.0285.00 ± 0.05
MMRGlobal pooling85.80 ± 0.0391.84 ± 0.0283.88 ± 0.05
Raw regional features86.20 ± 0.0391.96 ± 0.0284.14 ± 0.04
Graph-free MLP86.27 ± 0.0392.09 ± 0.0284.38 ± 0.04
Graph encoder (full GRef)87.94 ± 0.0093.08 ± 0.0185.34 ± 0.03

For each seed, we averaged the six datasets with equal weight. The table reports the mean ± sample standard deviation across three seeds. Values shown as ±0.00 are due to rounding.

Q1: How does GRef compare with raw regional features and MMR?

With the same regional inputs and MMR rule, mean image-level AUROC was 86.20% for raw features, 86.27% for the graph-free MLP and 87.94% for the graph encoder. Learning regional features alone added 0.07 points, while the graph encoder gained 1.74 points over raw features. The graph encoder was the strongest of these three regional representations under the same MMR rule.

Q2: What does graph encoding add over a graph-free MLP with MMR fixed?

The MLP and graph encoder used the same regional inputs, pretraining objectives and training budget. Modeling spatial adjacency and hierarchical containment gained 1.67 percentage points in mean image-level AUROC with MMR and 1.50 points with Top-K. The graph encoder outperformed the MLP under both selection rules.

Q3: What does MMR add over Top-K with the graph encoder fixed?

With the same graph encoder, MMR (λ = 0.7) gained 0.42 percentage points in image-level AUROC over relevance-only Top-K (λ = 1). Pixel-level AUROC and AUPRO also rose by about 0.17 and 0.34 points, respectively. Across all eight configurations in the table, full GRef, combining the graph encoder and MMR, ranked first in image-level AUROC, pixel-level AUROC and AUPRO.

Reviewer_p9akOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRef, a reference selection method that improves existing defect detectors by selecting suitable references for a query image from normal images of a new product. A GIN and MLP encoder captures spatial and hierarchical relationships among regional features. The pretrained encoder retrieves candidate references, after which maximum marginal relevance (MMR) selects a relevant yet nonredundant reference set. The selector can be integrated into existing detectors as a plug-in or used during detector training. Across 6 datasets, the proposed method achieves an average image-level AUROC of 87.94%, rising to 93.70% when integrated with UniVAD.

Critical strengths
  • The paper presents a practical method that can improve existing defect detectors using K-shot of labeled data.
  • The evaluation compares Fixed, Random, Similarity, CPR, and GRef under matched conditions across 4 detector backbones, with the detector held fixed within each comparison. The plug-in and training-based evaluations demonstrate the method's utility across multiple detectors.
  • Reference selection yields substantial average improvements when used as a plug-in without additional training. The effects of detector retraining are reported separately.
Critical weaknesses
  • The analysis does not sufficiently isolate the contributions of the key components. It remains unclear which components account for the performance gains of the proposed method.
  • The empirical results do not clearly demonstrate how the theoretical analysis explains the method’s performance. Theorem 1 provides an upper bound on the expected error for new products drawn from the same task distribution. However, the paper does not examine whether encoder training improves the selection margin or whether greater selection stability translates into better detection performance. The authors also acknowledge that an independently pretrained selector may not align with the detector's requirements. These issues call for empirical verification that the proposed training procedure satisfies the conditions under which the theoretical bound is sufficiently tight.
Questions for authors

Clarifying the following points would help assess the methodological contributions and interpret the results.

  • Performance contributions: How does performance change when the relational representation, GIN/MLP modules, and MMR are individually ablated? Which component contributes the most? Could you compare these components with simpler alternatives, such as global pooling, an MLP-only encoder, and top-k selection, to isolate their contributions?

  • Connection between theory and experiments:

    • Does encoder training increase the score margin between selected and rejected candidates?
    • Is greater stability in GRef's reference selection associated with better detection performance on new products?
    • When empirical quantities are substituted into Theorem 1, is the resulting error bound sufficiently tight to meaningfully explain the observed performance?
2 comments
AuthorsOpenReview ↗

Response to Reviewer p9ak (Part I)

Dear Reviewer p9ak,

Thank you for noting that GRef works with several detectors. Our reply took longer because we ran the encoder ablations and compared selection margins and detection performance before and after training.

Response to Weakness 1 and Question 1 (contributions of the encoder and MMR)

Section 4.3 summarizes the component gains. Table 4 in Appendix E.1 reports all eight configurations.

Replacing the graph-free MLP with the graph encoder raised six-dataset mean image-level AUROC from 86.27% to 87.94% with MMR, a gain of 1.67 percentage points. The same replacement gained 1.50 points with Top-K. Of the three steps from global pooling to regional pooling, graph-free encoding and graph message passing, adding message passing gave the largest gain under both selection rules.

We paired each of four representations (global pooling, raw regional features, an MLP-only encoder, and the full graph encoder) with both Top-K and MMR. Within each seed, all variants used the same ResAD++ weights, normal reference pool, and test set, with M = 32 candidates and K = 8 references. The two learned encoders used the same regional inputs, pretraining objectives, and training budget.

Selection ruleRepresentationImage-level AUROC (%)Pixel-level AUROC (%)AUPRO (%)
Top-KGlobal pooling85.55 ± 0.0591.80 ± 0.0283.82 ± 0.06
Raw regional features85.95 ± 0.0591.90 ± 0.0284.03 ± 0.06
Graph-free MLP86.02 ± 0.0591.99 ± 0.0284.17 ± 0.05
Graph encoder87.52 ± 0.0492.91 ± 0.0285.00 ± 0.05
MMRGlobal pooling85.80 ± 0.0391.84 ± 0.0283.88 ± 0.05
Raw regional features86.20 ± 0.0391.96 ± 0.0284.14 ± 0.04
Graph-free MLP86.27 ± 0.0392.09 ± 0.0284.38 ± 0.04
Graph encoder (full GRef)87.94 ± 0.0093.08 ± 0.0185.34 ± 0.03

For each seed, we averaged the six datasets with equal weight. The table gives the mean ± sample standard deviation across three seeds. Values shown as ±0.00 are due to rounding.

Q1: Which change in the image representation gives the largest gain?

With MMR fixed, replacing global pooling with raw multiscale regional features gained 0.40 points in mean image-level AUROC. Training a graph-free MLP on the same inputs added 0.07 points, and adding GIN message passing gained a further 1.67 points. Encoding adjacency and containment between regions gave the largest gain, exceeding the first two steps combined.

Q2: What does MMR add when the representation is held fixed?

MMR improved mean image-level AUROC, pixel-level AUROC and AUPRO over Top-K for all four representations. The strongest alternative was the same graph encoder with relevance-only Top-K (λ = 1). Switching to MMR (λ = 0.7) added 0.42, 0.17 and 0.34 percentage points in image-level AUROC, pixel-level AUROC and AUPRO, bringing the scores to 87.94%, 93.08% and 85.34%. Full GRef ranked first on all three metrics among the eight configurations. Reducing redundancy improved both detection and localization even with the strongest encoder.

AuthorsOpenReview ↗

Response to Reviewer p9ak (Part II)

Response to Weakness 2 and Question 2 (selection stability and the generalization bound)

Section 3.4 discusses the bound. Appendix H.7 reports margins, estimated reference change rates, and detection performance.

After training, the mean full selection margin across the six target datasets reached 2.07 times its value at initialization. With ResAD++ weights fixed, mean image-level AUROC rose from 86.89% to 87.94%. We compared the same RegionGIN architecture at initialization and after training, using the same test images, normal reference pools and reference budget. Each seed included 7,851 queries. We averaged the six datasets equally within each seed and report the mean ± sample standard deviation across three seeds.

Q1: Does encoder training increase the full selection margin?

Encoder training increases the full selection margin. For each query, we computed Γ as defined in the paper. We took the gap between the retrieval scores at ranks 32 and 33 and the gap between the best and second-best candidates at each of the eight MMR steps. Γ is the smallest of these nine gaps.

Full selection marginAt initializationAfter training
Mean margin (×10⁻⁶)7.243 ± 0.18814.985 ± 0.635
Mean of dataset medians (×10⁻⁶)4.076 ± 0.1228.438 ± 0.406
Mean of dataset 10th percentiles (×10⁻⁶)1.054 ± 0.0482.183 ± 0.128
Queries with Γ ≤ 10⁻⁶ (%)15.44 ± 0.354.69 ± 0.29

The mean margin increased on all six datasets. At the fixed threshold Γ ≤ 10⁻⁶, the fraction of small-margin queries fell from 15.44% to 4.69%. Training thus reduced the share of queries near a retrieval or MMR decision boundary by about 70%. These are the cases counted by the small-margin term in Theorem 1.

Q2: Is greater selection stability associated with better detection performance?

At the same feature perturbation magnitude, ε = 5 × 10⁻⁷, the response model estimated that the reference change rate fell from 11.03% to 1.75%, predicting more stable selection after training. We then evaluated the fixed ResAD++ on unperturbed images using image-level AUROC and mean pixel L₁ error. The latter is the detection loss in the theorem.

Mean over six datasetsAt initializationAfter training
Image AUROC(%)86.89 ± 0.0287.94 ± 0.00
Mean pixel L₁ error0.509792 ± 0.0080240.506708 ± 0.007777

Values shown as ±0.00 are due to rounding.

After training, the low- and high-margin groups had similar mean pixel L₁ errors on unperturbed images. Comparing the selector before and after training, mean image-level AUROC for the fixed ResAD++ rose by 1.05 percentage points, and mean pixel L₁ error fell from 0.509792 to 0.506708. The same detector performed better with the trained selector, and the response model predicted more stable reference selection under perturbation.

Q3: Is the bound in Theorem 1 tight enough to explain the observed performance?

Theorem 1 gives a quantitative explanation of reference selection stability. By Lemma 4, if changing the encoder shifts every descriptor by at most γ/2, then every query with margin Γ > γ retains the same Top-M candidate set and the same MMR selection order. At γ = 10⁻⁶, training reduced the fraction of small-margin queries from 15.44% to 4.69%. Equivalently, for the same bound on descriptor shifts, the fraction of queries covered by the stability guarantee rose from 84.56% to 95.31%, averaged equally across the six target datasets.

Training therefore left fewer queries close to a reference selection boundary. For the full bound on pixel L₁ risk, this fraction alone does not determine how close the bound is to the observed risk, since the bound also includes empirical risk on historical tasks and a complexity term. A similar margin–complexity trade-off appears in the graph classification bound of Li et al. Li et al., ICLR 2025, Theorem 5.1

Reviewer_k5cjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GRef addresses reference selection for industrial anomaly detection on unseen products, selecting K normal images from a larger pool for each query. A GIN encodes spatial adjacency and containment among multiscale image regions; the resulting embeddings retrieve M candidates, from which maximum marginal relevance (MMR) selects a relevant but less redundant subset. The encoder is pretrained on source normal images using dropout consistency and similarity preservation, then frozen. The selector supports both detector training and inference with an existing detector, without target-side parameter updates. Across six datasets, the reported mean image AUROC is 87.94% for ResAD++ trained with GRef and 93.70% for UniVAD with GRef. The paper also reports localization and runtime results and derives a generalization bound involving selection margins.

Critical strengths
  1. The selector addresses a practical constraint and works with several detectors. A deployed system may have many verified normal images but limited time to compare each query against them. GRef handles this choice through the detector's reference input, without requiring target-side parameter updates. Its compatibility with several existing detectors makes the contribution useful beyond a single architecture.
  2. The fixed-detector comparisons provide convincing evidence of improved reference selection. With the same pool and K = 8, GRef improves mean image AUROC over CPR by 1.47, 1.20, and 1.33 percentage points for PatchCore, RegAD, and ResAD++, respectively. Tables 6 and 7 also show average localization improvements. Holding the detector fixed makes these gains directly attributable to the choice of references.
Critical weaknesses
  1. The experiments do not isolate the contributions of graph encoding and redundancy reduction. Relative to Similarity, GRef changes the spatial features, encoder, pretraining objective, and selection rule together. The parameter studies do not resolve this confounding: there is no graph-free model using the same regional features and no pure relevance result with lambda = 1. It therefore remains unclear whether graph reasoning is necessary or whether multiscale features and ordinary MMR would suffice. Controlled comparisons with these alternatives, followed by separate edge and loss ablations, are needed to support the central design claims.
  2. Unclear run protocols and baseline settings weaken the evidence for some reported gains. UniVAD improves from 93.61% with Similarity or CPR to 93.70% with GRef. Training with GRef raises ResAD++ from 87.77% to 87.94% but lowers RegAD from 82.76% to 82.42%. The paper does not specify whether the three runs independently retrain both networks, so the reported deviations do not establish whether these small differences are reliable. Independent seeds and paired uncertainty estimates are needed.
Questions for authors
  1. Can you provide the controlled comparisons in Weakness 1, keeping the pool, K, and detector weights fixed? Please also remove each edge type and each term of Equation (3) separately.
  2. What explains the AnomalyCLIP discrepancy? Please give the checkpoint, source training setup, input resolution, preprocessing, and image-level scoring rule, and check the result against the official implementation under matched settings.
  3. How does the implemented model satisfy Equation (9) and the descriptor nondegeneracy assumption? What is counted as an independent task in m? Reporting the selection-margin distribution and its relation to detection error would help explain the practical relevance of the bound.
6 comments
AuthorsOpenReview ↗

Response to Reviewer k5cj (Part I)

Dear Reviewer k5cj,

Thank you for your comments on the fixed-detector comparisons and for your patience with our delayed reply. We took the extra time to run the component ablations and ten-seed comparisons and to evaluate the released AnomalyCLIP checkpoints.

Response to Weakness 1 and Question 1 (controlled component ablations)

Section 4.3 summarizes the graph encoder and MMR ablations. Appendix E.1 gives the full results in Tables 4 and 5.

With the same regional inputs and MMR, the graph encoder improved mean image-level AUROC by 1.67 percentage points over the graph-free MLP. With the graph encoder fixed, MMR gained 0.42 points over relevance-only Top-K. These comparisons show the gains from using the graph encoder and from reducing reference redundancy, respectively. Within each seed, all variants shared the ResAD++ weights, normal pool and test set. They selected 8 references from 32 candidates.

Q1: What does graph encoding add over simpler regional representations?

We pair each of four representations with both Top-K and MMR. The graph-free MLP and the graph encoder use the same multiscale regional features, source training data, pretraining objective and training budget.

Selection ruleRepresentationImage-level AUROC (%)Pixel-level AUROC (%)AUPRO (%)
Top-KGlobal pooling85.55 ± 0.0591.80 ± 0.0283.82 ± 0.06
Raw regional features85.95 ± 0.0591.90 ± 0.0284.03 ± 0.06
Graph-free MLP86.02 ± 0.0591.99 ± 0.0284.17 ± 0.05
Graph encoder87.52 ± 0.0492.91 ± 0.0285.00 ± 0.05
MMRGlobal pooling85.80 ± 0.0391.84 ± 0.0283.88 ± 0.05
Raw regional features86.20 ± 0.0391.96 ± 0.0284.14 ± 0.04
Graph-free MLP86.27 ± 0.0392.09 ± 0.0284.38 ± 0.04
Graph encoder (full GRef)87.94 ± 0.0093.08 ± 0.0185.34 ± 0.03

With MMR fixed, the graph encoder gained 1.74 percentage points in mean image-level AUROC over raw regional features. Even with relevance-only Top-K, it gained 1.50 points over the graph-free MLP. The graph encoder ranked first among the four representations under either selection rule. Across all eight combinations, full GRef achieved the highest image-level AUROC, pixel-level AUROC and AUPRO.

Q2: How does relevance-only selection at λ = 1 compare with MMR?

The parameter λ controls the relative weights of query relevance and the redundancy penalty. The score for a candidate reference r is

score⁡(r)=λs(q,r)−(1−λ)max⁡r′∈Ss(r,r′)\operatorname{score}(r) = \lambda s(q,r) - (1-\lambda)\max_{r'\in S}s(r,r')

where s(q, r) measures the relevance of r to the query q, and S is the set of references already selected. At λ = 1, the redundancy term has zero weight, so the candidates are ranked entirely by relevance. At λ = 0.7, the redundancy term has weight 0.3. The first reference is still selected by relevance. In later steps, candidates similar to an already selected reference receive a penalty, favoring relevant candidates that add less redundant information.

Selection ruleλImage-level AUROC (%)Pixel-level AUROC (%)AUPRO (%)
Top-K (relevance only)1.087.52 ± 0.0492.91 ± 0.0285.00 ± 0.05
MMR (full GRef)0.787.94 ± 0.0093.08 ± 0.0185.34 ± 0.03

With the graph encoder fixed, the redundancy penalty improved image-level AUROC, pixel-level AUROC and AUPRO by about 0.42, 0.17 and 0.34 percentage points, respectively. With the same eight reference slots, GRef’s redundancy-aware selection outperformed relevance-only Top-K on all three detection and localization metrics.

AuthorsOpenReview ↗

Response to Reviewer k5cj (Part I, continued)

Response to Weakness 1 and Question 1 (continued)

Table 5 reports the loss ablations.

Q3: What does each loss term in Eq. (3) contribute?

We keep the graph encoder architecture and MMR rule fixed, remove one pretraining loss term at a time, and retrain the encoder.

Training objectiveImage-level AUROC (%)Pixel-level AUROC (%)AUPRO (%)
Both loss terms87.94 ± 0.0093.08 ± 0.0185.34 ± 0.03
Without dropout consistency87.55 ± 0.0592.95 ± 0.0184.90 ± 0.05
Without similarity preservation87.40 ± 0.0592.90 ± 0.0284.75 ± 0.07

Removing dropout consistency lowered mean image-level AUROC by 0.39 percentage points. Removing similarity preservation lowered it by 0.54 points. Both localization scores also dropped in each ablation. Training with both losses gave the best detection and localization scores of the three objectives tested.

Q4: What do the two edge types represent, and what does the graph-free comparison test?

Spatial adjacency connects neighboring regions at the same scale, and containment connects parent and child regions across scales. All edges share the same rule for computing edge weights and the same message-passing parameters. Section 3.1 now describes these two connectivity patterns and the parameters they share. Removing one pattern tests its contribution while messages still pass along the remaining edges. Removing all inter-region edges leaves each region to update from its own features. Our ablation therefore tests the joint contribution of inter-region message passing, using a graph-free MLP with the same regional inputs. The graph encoder gained 1.50 percentage points in mean image-level AUROC over the MLP with Top-K and 1.67 points with MMR.

For all three tables, we first average the six datasets equally within each seed, then report the mean ± sample standard deviation across three seeds. Values shown as ±0.00 are due to rounding.

AuthorsOpenReview ↗

Response to Reviewer k5cj (Part II)

Response to Weakness 2 (run protocol and ten-seed results)

Appendix D gives the training protocol for each seed, and Table 12 in Appendix G.1 reports the ten-seed results.

With UniVAD fixed, GRef achieved 93.74% mean image-level AUROC over ten seeds, compared with 93.60% for Similarity and 93.58% for CPR. GRef therefore retained its advantage over both selectors on the strongest detector in our study.

Q1: What is trained or held fixed for each seed?

For each seed, we train a GRef selector from scratch and reuse it in all GRef comparisons for that seed and target dataset. We also train the standard RegAD and ResAD++ detectors independently for each seed. Within a seed, the five inference-time reference strategies share the corresponding detector weights, normal pool, reference budget and test set. UniVAD uses the same pretrained weights across all seeds. For each seed, Similarity, CPR and GRef each select eight references according to their own rule, and the same detector evaluates all three. The standard deviation for GRef includes variation from retraining the selector.

Q2: How do we isolate the effect of using GRef during detector training?

For each seed and each detector (RegAD or ResAD++), the GRef and GRef† runs start from the same detector initialization. They use the same training images, data order, augmentation rules, optimizer settings, number of training steps and checkpoint selection rule. Only the choice of training references differs. At test time, the two runs use the same GRef selector, selected references and queries. The difference GRef† − GRef therefore measures the effect of changing the training references when both runs already use GRef at inference.

Q3: What do the ten-seed comparisons show?

The table gives the mean and sample standard deviation of image-level AUROC across ten seeds. Within each seed, we weight the six datasets equally. GRef denotes selection at inference only. GRef† uses the selector during both training and inference.

DetectorReference selectionMean image-level AUROC (%)
RegADGRef82.79 ± 0.27
RegADGRef†82.86 ± 0.30
ResAD++GRef87.81 ± 0.12
ResAD++GRef†88.05 ± 0.13
UniVADSimilarity93.60 ± 0.08
UniVADCPR93.58 ± 0.11
UniVADGRef93.74 ± 0.03

Within each seed, we compared GRef with Similarity and with CPR separately on UniVAD, and GRef† (training and inference) with GRef (inference only) separately on RegAD and ResAD++, then averaged the ten differences for each comparison. With UniVAD fixed, GRef gained 0.14 percentage points over Similarity and 0.16 over CPR on average. Across the ten seeds, using GRef during training as well as inference raised mean image-level AUROC for ResAD++ by 0.24 percentage points relative to inference alone.

In the three runs reported in Table 2, using GRef during training lowered RegAD’s mean image-level AUROC by 0.34 percentage points relative to using GRef at inference alone. Over ten seeds, the same comparison gave a small average gain of 0.07 points. This change in direction is consistent with seed-to-seed variation.

In the fixed-detector comparisons in Table 2, replacing Fixed with GRef increased mean image-level AUROC by 6.90 points for RegAD and 5.92 for ResAD++. Against CPR, GRef gained 1.20 and 1.33 points, respectively. Both detectors gained most from reference selection at inference. Across ten seeds, both detectors showed a small additional mean gain when GRef was also used during training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.