acceptodds
Withdrawn from ICLR 2027

Tri-Mem: A Triple-Memory Framework for Online Industrial Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Industrial anomaly detection in practical inspection systems often relies on only a few normal reference images and must accommodate appearance variations encountered during testing. Most existing memory-based few-shot methods use a fixed repository of normal features, which limits their ability to incorporate information from the test stream and to use anomaly-related evidence. We propose Tri-Mem, a training-free framework for online few-shot anomaly detection that consists of an independent frozen core memory, an adaptive normal memory, and an anomaly memory. The core memory preserves support-derived normal prototypes, while the other two memories are updated periodically with rank-selected test features and compressed independently. Tri-Mem further combines multi-layer representations from a frozen DINOv3 encoder with support-time illumination augmentation and capacity-triggered coreset compression of the adaptive memories. We evaluate the proposed method under 1-, 2-, 4-, 8-, and full-shot protocols on four industrial anomaly detection benchmarks. On MVTec AD 2, Tri-Mem yields consistently strong pixel-level results across the evaluated few-shot settings. In the full-shot setting, it obtains 91.2% P-AUROC, 64.6% P-AUPRO, and 72.6% I-AUROC. Competitive results are also observed on MVTec AD, MPDD, and BTAD, particularly for pixel-level localization. Ablation experiments indicate that the three memory components provide complementary information: compared with the static core-memory baseline, the complete configuration increases the average P-AUPRO and I-AUROC on MVTec AD 2 by 4.6 and 3.2 percentage points, respectively.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Industrial defect detection, Few-shot anomaly detection, Online anomaly detection, Dynamic memory bank

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_3owDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Tri-Mem is an online few-shot method for industrial anomaly detection. It keeps the initial normal support features fixed and maintains two additional memories using low- and high-scoring patches from previous test images. Using frozen DINOv3 features, it combines distances to normal references with a bounded score from the anomaly memory. The authors evaluate the method on four industrial datasets and provide component ablations. On MVTec AD 2, the full method improves average pixel AUPRO by 4.6 percentage points and image AUROC by 3.2 points over the fixed-memory baseline across four support budgets.

Critical strengths
  1. Support-image augmentation helps reduce sensitivity to lighting changes when normal references are scarce. The three-memory design preserves the initial support features while adding normal and suspicious patterns from the test stream, allowing adaptation without replacing the original references.
  2. The method provides strong region-level localization results. It achieves the highest reported pixel AUPRO across all four few-shot budgets on MVTec AD 2 and at 2-, 4-, and 8-shot on BTAD. The memory and augmentation ablations provide evidence for the benefits of the main components.
Critical weaknesses
  1. The reliability of online updates is not sufficiently evaluated. The method selects normal and suspicious patches using its own scores, so selection errors can affect later predictions. However, evaluation uses only one fixed test-stream order, with no reported variation across repeated runs. Candidate purity and error accumulation are left for future work. The reported gains therefore need validation across different stream orders.
  2. Comparisons with closely related online methods are incomplete. The paper discusses few-shot online anomaly detection by Wei et al. but provides no quantitative comparison. Several online baselines are implemented by the authors, yet their selection and update rules are not sufficiently described. This limits the assessment of the gains over existing online approaches.
Questions for authors
  1. Do the gains remain consistent when the test images arrive in a different order? Please report the mean and standard deviation over several test orders, comparing the full method with the frozen core and the core plus normal memory.
  2. Can the authors add comparisons with other online anomaly detection methods under the same few-shot setting and briefly describe any changes made to the baseline methods?
  3. The joint-compression ablation allows the initial support features to be removed and also reduces the total normal-memory capacity. Can the authors keep the total capacity the same and compare keeping these features fixed versus allowing them to be removed? This would clarify the benefit of protecting the initial references.
  4. The augmentation ablation removes brightness changes and RGB shifts together. Since color changes can themselves be defects, can the authors test RGB shifts separately and report their effect on detecting color defects?
Reviewer_Mb5COpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies online few-shot industrial anomaly detection. In the setting assumed by the authors, only a few normal support images are available, and object appearance (e.g., illumination, reflections) changes over an unlabeled test stream. In this environment, the authors identify two failure modes: fixed-memory methods such as PatchCore may mistake benign appearance changes for defects, whereas naively updating the memory with unlabeled test patches can propagate errors to later images.

Tri-Mem is a training-free, retrieval-based detector with three memories built in a shared feature space, obtained from multiple intermediate-layer features of a frozen DINOv3 ViT-B/16 and a support-fitted PCA whitening.

  • Core Memory MC\mathcal{M}_C (Section 3.1): support descriptors from the original support images and several illumination-augmented views, deduplicated by order-preserving cosine filtering and kept frozen during inference.
  • Normal Memory MN\mathcal{M}_N and Anomaly Memory MA\mathcal{M}_A (Section 3.2): after each test image is scored, its lowest-scoring and highest-scoring patches are stored in separate caches, which are periodically merged into the corresponding memories after deduplication and K-center compression. Patch scores combine the kNN ℓ1+ℓ2\ell_1+\ell_2 distance to MC∪MN\mathcal{M}_C \cup \mathcal{M}_N with a bounded proximity term to MA\mathcal{M}_A.

The key claims are:

  1. Separating a frozen support core from two independently bounded adaptive memories allows online evidence to change later scores without displacing the support evidence.
  2. Score-before-update retrieval in a fixed, support-fitted space yields a causal online detector without test-time gradient updates.
  3. Tri-Mem achieves the highest P-AUROC and P-AUPRO on MVTec AD 2 across all few-shot budgets, and the full configuration improves average P-AUPRO and I-AUROC over the frozen core.

Experiments use MVTec AD 2 (public test split) as the main benchmark and MVTec AD, MPDD, and BTAD as additional benchmarks, comparing six CLIP-based and non-CLIP baselines in offline and author-implemented online variants. P-AUROC, P-AUPRO, and I-AUROC are reported from single runs.

Critical strengths

S1 (Quality): The localization results on MVTec AD 2 are strong.

  • Tri-Mem ranks first in P-AUROC and P-AUPRO across all four few-shot budgets.

S2 (Clarity/Reproducibility): The online protocol and implementation are specified precisely.

  • The score-before-update order and flush schedule (Section 3.2) are clearly defined, so it is unambiguous which memory state each image is scored against.
  • The illumination operators are described in reproducible detail, down to their distributions and ranges (Appendix A.1).
Critical weaknesses

W1 (Originality): The method is largely a combination of existing components.

  • Its main components all come from prior work:
    • DINO patch-feature kNN for few-shot AD: AnomalyDINO
    • Layer selection: Dinomaly
    • Coreset-based memory: PatchCore
    • Whitening with support statistics
  • The problem setting is not new either. Initializing from few-shot normal data and updating with pseudo-labels from an unlabeled stream that mixes normal and anomalous samples was already formulated by K-NG (Wei et al., 2024). The paper cites this work but neither differentiates from it nor compares against it.
  • MuSc (ICLR 2024), which also exploits normal and abnormal cues in unlabeled test images, is not discussed.
  • The genuinely new element is therefore the three-way memory split with a bounded MA\mathcal{M}_A term, and this is precisely the part not sufficiently supported by the ablations (W2).

W2 (Quality): The ablations do not clearly show the contribution of each adaptive memory.

  • The full configuration improves P-AUPRO and I-AUROC over the core but not P-AUROC, and it is not explained which component this improvement comes from.
    • MC+MA\mathcal{M}_C+\mathcal{M}_A is identical to MC\mathcal{M}_C in almost all entries.
    • MC+MN\mathcal{M}_C+\mathcal{M}_N also does not appear to make a substantial standalone contribution. The improvement is small, and without variance estimates it cannot be distinguished from noise.
  • Nevertheless, combining all three memories produces a clear improvement. The paper does not explain why combining components that individually have little effect yields this gain, and the authors themselves acknowledge that the contribution of MA\mathcal{M}_A cannot be identified (Appendix C.4.2).

W3 (Quality): The capacity-control mechanisms appear inactive under the default settings.

  • Given the number of patches selected per image at the default resolution and the flush interval, the number of descriptors accumulated in each cache before a flush is only a fraction of the cache capacity. Trimming therefore appears never to occur, and no analysis of this is provided.
  • Exceeding the adaptive-memory capacities would require several hundred or more test images, even before deduplication. K-center compression is therefore also likely to trigger rarely or never.
  • As a result, the capacity control, which occupies a substantial part of the method description, may never have been exercised, and no occupancy statistics are reported to show otherwise.

W4 (Soundness): The method does not address the central problem raised in the introduction.

  • The introduction presents error propagation from wrongly selected unlabeled patches as the key challenge (Section 1).
  • However, normal retrieval uses the union of MC\mathcal{M}_C and MN\mathcal{M}_N (Eq. 5). Once a defective patch enters MN\mathcal{M}_N, the distance for similar defects in later images decreases, and freezing the core does not block this pathway.
  • No mechanism verifies candidate purity, and this is deferred to future work (Section 5).
  • The introduction also does not state which failure case MA\mathcal{M}_A is meant to resolve. As a result, the stated problems and the proposed modules do not correspond.

W5 (Quality): The baseline comparison is not reliable enough to support the state-of-the-art claim.

  • Some author-implemented online variants behave anomalously:
    • RareCLIP yields below-chance I-AUROC at some shots.
    • UniVAD's performance drops sharply when moved online (Table 1).
  • The paper does not state whether the baselines use the same backbone and resolution as Tri-Mem. A controlled baseline, such as PatchCore with DINOv3 features, is also missing.
  • There is no comparison with K-NG (Wei et al., 2024), which targets the same setting, nor with the closely related online extension of PatchCore based on incremental k-center selection (Ren et al., 2025).

W6 (Quality): The results lack statistical support.

  • All numbers come from single runs without seed variance, despite stochastic components such as augmentation sampling and K-center initialization.
  • Although the method is online, it is evaluated on a single fixed stream order, with no sensitivity analysis over permutations.

W7 (Quality): Core design choices are presented without justification.

  • (a) PCA: The only rationale given is placing all descriptors in a single space (Section 3.1). However, since the encoder is frozen, all descriptors are already compared in the same space even if the original high-dimensional features are used without PCA. No justification is therefore given for why PCA is needed at all.
  • (b) ℓ1+ℓ2\ell_1+\ell_2 distance:
    • There is no theoretical motivation.
    • In high dimensions, ℓ1\ell_1 is generally much larger than ℓ2\ell_2, so the score is effectively determined by ℓ1\ell_1.
    • Neighbors are selected by ℓ2\ell_2 but scored by ℓ1+ℓ2\ell_1+\ell_2, which is inconsistent.
  • (c) Core construction:
    • The core is built by order-dependent near-duplicate removal with a very high threshold, which can hardly be considered a selection of representative "prototypes."
    • It uses a different metric (cosine similarity) from retrieval.

W8 (Clarity): The figures do not match the text and are difficult to follow.

  • Figure 2 depicts the core as built by a "K-center Coreset," whereas the text uses cosine deduplication.
  • Figure 1 uses a "Training" label for a training-free method.
  • Figure 3 does not specify the shot count or whether each baseline is offline or online.

W9 (Clarity): The notation and organization require a thorough revision.

  • Many symbols are left undefined.
  • Several symbols are used before their definition or overloaded.
  • The Related Work section merely lists methods and does not explain how Tri-Mem differs from the closest online few-shot approaches.
Questions for authors

Q1 (W2): What does each main component of the proposed method actually contribute?

  • In the ablations, adding MN\mathcal{M}_N or MA\mathcal{M}_A alone yields little or no improvement, and a clear improvement appears only when all three memories are combined. How do you explain this result?
  • Please provide analyses showing that each component actually functions as intended, for example, how often each memory is used for retrieval during inference and how much the bounded anomaly term affects the final scores.
  • Are there any implementation differences between the full configuration and each ablated configuration other than the presence or absence of the corresponding memory?

Q2 (W1): What is the methodological novelty of Tri-Mem compared with existing methods?

  • Beyond combining existing components (DINO patch-feature kNN as in AnomalyDINO, the layer selection of Dinomaly, coreset-based memory as in PatchCore, and whitening with support statistics), what do you consider the core methodological contribution of Tri-Mem?
  • K-NG (Wei et al., 2024) addresses the same setting of initializing from few-shot normal data and updating with pseudo-labels from an unlabeled stream, and Ren et al. (2025) extend PatchCore to online learning on incoming normal samples via incremental k-center selection. Apart from the use of additional memory banks, how does Tri-Mem differ from these methods, and what concrete advantages does it offer over them?
  • How does Tri-Mem differ from MuSc (ICLR 2024), which also exploits normal and abnormal cues in unlabeled test images?

Q3 (W3): Were the capacity mechanisms ever activated?

  • Please report the test-stream length for each dataset and category.
  • How many times were trimming and K-center compression actually triggered?
  • How did the sizes of MN\mathcal{M}_N and MA\mathcal{M}_A evolve across flushes?

Q4 (W4): How pure are the normal candidates?

  • Measured post hoc with the test ground truth, what fraction of the patches admitted to MN\mathcal{M}_N lie inside defective regions?
  • Is this fraction associated with false negatives on later images?

Q5 (W5): Are the baselines comparable?

  • Which backbone and input resolution does each baseline use?
  • Could you add comparisons with PatchCore and AnomalyDINO using the same DINOv3 backbone, as well as with K-NG?

Q6 (W6): How robust are the results?

  • How was the stream order determined?
  • Could you report means and standard deviations over multiple stream permutations and seeds?

Q7 (W7): What is the rationale for using PCA instead of the original high-dimensional feature space?

  • How was the PCA dimension chosen?
  • Could you provide a comparison with the original feature space without PCA across all shot settings?
Reviewer_G72COpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an online few-shot anomaly detection method designed to adapt to changes at test time. It preserves support-derived descriptors in a frozen core memory and updates separate normal and anomaly memories from an unlabeled test stream. The paper presents this as a triple-memory framework.

Critical strengths

None.

Critical weaknesses
  1. The central design choice, three memory banks, is neither motivated nor justified.
  • The introduction is short and only states the high-level design. It does not explain the benefit of splitting the method into three memory banks. For example, why not use two, one for normal and one for abnormal? Why is one of the memories fixed, and why are both the normal and the abnormal memories trained?
  • The terms "Foundational Knowledge" and "Contrastive Knowledge" could partly answer these questions if they were elaborated, but they appear only inside Fig. 2 and are never used again, not even in the caption.
  • I also do not see a good ablation study in the main paper that justifies using three memory banks or choosing which ones are trained and which are fixed. Without this, I cannot tell which parts of the design matter.
  1. The method description lacks goals, and the figures and notation are hard to follow.
  • Secs. 3.1 and 3.2 start with implementation details without a high-level idea or objective. For the frozen core, it is unclear what the authors want to achieve, why PCA is used, why augmented views are needed, or why descriptors are retained only when their cosine similarity to all previously retained descriptors is at most τ. These details matter little when the reader does not know the goal.
  • Notation is unclear. E2,E4,E6,E8E_2, E_4, E_6, E_8 appear in Eq. 1 without definition (I can only assume they are encoders), as do HfH_f and WfW_f. The image index tt is never linked to the subscript of the test stream {I…test}\{I^{\text{test}}_{\dots}\}.
  • Fig. 2 is not walked through. L122 says "As shown in Figure 2, Stage I ...", but neither Stage I nor Stage II is annotated in the figure, and the caption does not say the pipeline is two-stage.
  • At L124, "adaptive" is used loosely for both MAM_A and MNM_N, while Fig. 1 uses "adaptive" for MNM_N and "Contrastive" for MAM_A.
Questions for authors
  1. L14: adding "memory" to the text "frozen core" would help highlight the core idea in the title, TRIPLE-MEMORY FRAMEWORK.
  2. L46: MNM_N and MAM_A should be paired more closely in the writing with the text "normal memory" and "anomaly memory".
  3. Use "adaptive" consistently for MAM_A and MNM_N at L124 and in Fig. 1, and add a walkthrough of Fig. 2 to the main text and caption.

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

Related papers

Discussion (0)

Sign in to comment.