acceptodds
Withdrawn from ICLR 2027

Hyper-FSAD: Training-Free and Language-Free Few-Shot Anomaly Detection via Sparse Hyper Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-shot anomaly detection (FSAD) is particularly valuable when only a few normal images are available in a new target domain, while anomalous cases are rare, diverse, and difficult to enumerate in advance. However, existing methods often still require task-specific fitting or language prompts, and their patch-level retrieval commonly relies on brittle nearest-neighbor or fixed Top- rules. We propose **Hyper-FSAD**, a training-free and language-free framework that performs support-only inference with a frozen visual encoder. We formulate FSAD as support-only scoring in frozen feature space and establish the selective stability of sparse retrieval and sufficient conditions for normal–anomaly separation. Guided by this analysis, **Sparse Hyper Matching** uses **sparsemax** to adaptively select support patches for each query patch, exactly suppressing below-threshold distractors without a manually specified retrieval hyperparameter. **Dual-Branch Image Scoring** further combines local reconstruction discrepancies with support-conditioned global <CLS> deviation. Across four industrial and two medical benchmarks, Hyper-FSAD achieves the best overall performance across the 1/2/4-shot settings, while requiring only 52.6 ms per image and 0.89 GB GPU memory. The code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Few-Shot Anomaly Detection

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_pwf4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Hyper-FSAD, a training-free and language-free few-shot anomaly detection method based on frozen visual features. Its local branch uses sparsemax to establish sparse correspondences among normal support patches and reconstruct query patches, generating anomaly maps from reconstruction residuals. The global branch computes anomaly scores based on the distances between the <CLS> features of query images and normal support images. The paper conducts experiments on six industrial and medical datasets and multiple frozen backbones, and provides a conditional theoretical analysis of sparse matching.

Critical strengths
  1. The method requires neither target-task training nor language input and has a simple architecture. Experiments with three frozen backbones, six industrial and medical datasets, and five support set samplings demonstrate good applicability and stability.

  2. The paper provides a relatively comprehensive experimental evaluation. Ablation studies comparing Top-1, Top-p, softmax, entmax, and sparsemax under the same backbone directly examine the retrieval rules. Additional analyses are also provided for dual-branch fusion and layer selection.

  3. Sparsemax produces exact zero weights and an adaptive active support set, offering a clear, training-free trade-off between single nearest-neighbor matching and dense averaging.

Critical weaknesses
  1. The comparisons with existing methods use their respective original backbones, resolutions, and training protocols, while the results of the proposed method are obtained using the newer DINOv3 backbone. This makes it impossible to disentangle the contributions of sparsemax and the dual-branch design from differences in pretrained models.

  2. The image-level score uses max pooling over the local anomaly map, making it susceptible to domination by a single noisy patch. The global fusion weight is also fixed at 0.5, without sufficient cross-dataset calibration or hyperparameter analysis.

  3. Some few-shot methods that achieve better performance than the proposed method are not included in the comparisons, such as UniVAD (CVPR 2025).

Questions for authors

Could you provide additional performance comparisons with existing methods using the same pretrained backbone to better demonstrate the contributions of the proposed method?

Reviewer_Ea8oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Hyper-FSAD constructs a reference memory from a few normal images using a frozen visual encoder. It generates anomaly maps from local feature discrepancies after Sparsemax-weighted reconstruction and combines local and global CLS scores for image-level detection, without additional task-specific training or textual inputs. The paper provides theoretical analysis and evaluates the method on six industrial and medical datasets.

Critical strengths
  • S1: The method removes additional task-specific training and textual prompts, reducing adaptation requirements in few-shot deployment. The overall pipeline is easy to follow, and combining sparse local reconstruction with global normality evidence is reasonably motivated.

  • S2: The evaluation covers six industrial and medical datasets, multiple visual backbones, and different shot settings, with ablations of the major design choices. The reported results are generally competitive and demonstrate the practical potential of this simplified framework.

Critical weaknesses
  • W1: AnomalyDINO [1] already performs training-free, vision-only few-shot anomaly detection using frozen visual features and nearest-neighbor retrieval from normal patches. DictAS [2] already employs sparse dictionary lookup, weighted reconstruction, and cosine-discrepancy scoring; it also freezes the visual encoder but learns additional Q/K/V transformations through auxiliary training, without requiring retraining for new target classes. In contrast, Hyper-FSAD applies sparsemax directly to frozen patch embeddings, without learned Q/K/V transformations or auxiliary training, and incorporates support-conditioned CLS scoring. This offers useful simplification, but the principal methodological advance or new insight remains insufficiently articulated. The theory provides a useful conditional rationale for selective retrieval and local–global scoring, but its contribution to distinguishing Hyper-FSAD from closely related approaches remains unclear.

  • W2: Direct comparisons with [1] and [2] are missing, leaving the incremental value over the closest complete methods insufficiently established. Existing retrieval and layer-selection ablations support the local design, and Table 9 demonstrates a benefit from CLS fusion on MVTecAD. However, the effect of adding CLS to a nearest-neighbor baseline is not reported. The experiments also provide limited insight into cases where anomalous patches closely resemble normal reference patches, beyond the rejection of low-similarity candidates.

[1] Simon Damm, Mike Laszkiewicz, Johannes Lederer, and Asja Fischer. AnomalyDINO: Boosting Patch-Based Few-Shot Anomaly Detection with DINOv2. WACV, 2025.

[2] Zhen Qu, Xian Tao, Xinyi Gong, ShiChen Qu, Xiaopei Zhang, Xingang Wang, Fei Shen, Zhengtao Zhang, Mukesh Prasad, and Guiguang Ding. DictAS: A Framework for Class-Generalizable Few-Shot Anomaly Segmentation via Dictionary Lookup. ICCV, 2025.

Questions for authors
  • Q1: What is the principal technical increment over [1] and [2]? Please clarify how the core contribution should be positioned and how the theoretical and empirical evidence supports this positioning.

  • Q2: Could the authors compare the DINOv2 and CLIP versions of Hyper-FSAD with [1] and [2], respectively, on representative benchmarks, using matched normal support images and encoder checkpoints compatible with each baseline’s original protocol? Adding the nearest-neighbor baseline without CLS scoring would complete the existing 2×2 ablation on MVTecAD and help clarify the contribution of CLS fusion.

  • Q3: To clarify the mechanism’s scope, could the authors analyze representative cases from the existing benchmarks where annotated anomalous regions closely resemble normal reference patches, comparing nearest-neighbor and sparse reconstruction scoring? This would help explain when the method offers benefits beyond rejecting low-similarity candidates, as well as its limitations.

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

The paper proposes Hyper-FSAD, a training-free and language-free framework for few-shot anomaly detection using frozen visual encoders and a small set of normal support images. Its Sparse Hyper Matching module employs sparsemax to adaptively select relevant support patches for feature reconstruction, while Dual-Branch Image Scoring combines local reconstruction discrepancies with global semantic information. The paper further provides theoretical analysis of sparse retrieval stability and sufficient conditions for anomaly separation. Evaluated on four industrial and two medical datasets using CLIP, DINOv2, and DINOv3 backbones under 1/2/4-shot settings, Hyper-FSAD achieves competitive image- and pixel-level detection performance with low inference overhead. Additional ablations on matching strategies, feature layers, and scoring mechanisms support the effectiveness of the proposed design.

Critical strengths
  1. Hyper-FSAD adopts a training-free and language-free framework with frozen visual encoders, requiring only a few normal support images. Its sparsemax-based matching adaptively selects relevant support patches, while dual-branch scoring combines local reconstruction errors with global semantic evidence. Same-backbone ablations demonstrate improved anomaly localization, and the DINOv3 implementation achieves 52.6 ms inference time with only 0.89 GB GPU memory , supporting its practical deployment potential.
  2. The paper provides extensive evaluations across four industrial and two medical datasets, three visual backbones (CLIP, DINOv2, and DINOv3), and 1/2/4-shot settings. Five repeated runs with reported standard deviations, additional evaluation metrics, and ablations on matching strategies, feature layers, and scoring mechanisms provide evidence of the method's effectiveness and applicability across different settings.
Critical weaknesses
  1. Lacks innovation; Hyper-FSAD mainly combines existing techniques, including sparsemax, feature reconstruction, and local-global score fusion. Sparse Hyper Matching is essentially a sparse weighted reconstruction mechanism, and its distinction from existing retrieval-based anomaly detection methods remains unclear. Most individual components are already well established.
  2. The theoretical contribution is not sufficiently convincing. The sparse retrieval guarantees largely follow known properties of sparsemax, while the anomaly separation results rely on strong assumptions about normal coverage and anomaly margins. The authors do not empirically verify whether these assumptions hold in practice or demonstrate a clear theoretical advantage over alternative retrieval strategies. Consequently, the theoretical analysis does not sufficiently justify the superiority of the proposed method.
  3. Insufficient comparisons with relevant methods. The softmax baseline lacks temperature tuning, Top-p is evaluated with only a single fixed ratio, and closely related methods such as AnomalyDINO are omitted. Furthermore, the default DINOv3 model achieves lower average P-AUROC than SubspaceAD across all three shot settings. Despite extensive experiments, the paper does not convincingly establish the superiority of its proposed method over existing approaches, particularly under matched-backbone comparisons.
Questions for authors

Q1.Comparisons with relevant training-free methods are insufficient. In particular, AnomalyDINO, which is closely related to frozen-feature patch matching, is not included in the evaluation. Could the authors provide additional comparisons with relevant training-free approaches?

Q2.The comparison between sparsemax and alternative matching strategies is not sufficiently convincing. In Table 2, softmax is evaluated without temperature tuning, while Top-p uses only a fixed 1% ratio. Could the authors evaluate these methods under different parameter settings to demonstrate that the improvements genuinely come from sparsemax rather than suboptimal baseline configurations?

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

Related papers

Discussion (0)

Sign in to comment.