acceptodds
Withdrawn from ICLR 2027

LACToS: Local Anomaly Detection via Conditional Typicality of Superpixels

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Local anomaly detection is the task of identifying spatially confined anomalous or out-of-distribution (OOD) regions within an otherwise in-distribution (ID) image. Most existing methods are designed for a specific context, with a strong focus on industrial defect detection, where they achieve excellent performance on well-established benchmarks. However, we observe that these methods can exhibit significant performance drops in other contexts, especially when the inlier distribution is multimodal. To account for this inlier complexity, we propose LACToS, a principled local OOD detection method based on statistical typicality and made tractable through two innovations: (i) abstracting pixels into superpixels and (ii) using a computationally efficient proxy to estimate typicality. We evaluate LACToS across three diverse applications spanning medical imaging, human faces, and X-ray security screening. LACToS has more robust performance across these benchmarks compared to prior methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Out-of-distribution detection, diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ac4kOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

LACToS addresses local anomaly detection through the conditional typicality of superpixels. It uses an inlier diffusion model to generate shared denoising samples that approximate the otherwise expensive superpixel-specific inpainting posteriors. Per-superpixel histograms of pixel values and compressed features support anomaly scores based on deviations from the expected negative log-likelihood. The method is evaluated on medical imaging, human faces, and X-ray security screening together with toy experiments on inlier multimodality.

Critical strengths

The paper identifies a meaningful limitation of reconstruction-based anomaly detection. Collapsing plausible reconstructions into a point estimate can obscure multimodal inlier distributions. The proposed typicality score offers an interesting statistical perspective. Sharing denoising samples across superpixels makes its computation more tractable. The toy experiments directly examine the motivating problem of inlier multimodality. The evaluation also covers several application domains beyond industrial defect detection.

Critical weaknesses
  1. My main concern is the evaluation protocol. Appendix E.3 states that all LACToS hyperparameters are selected by maximizing AUROC on 20 test images. The selected configuration is then evaluated on the whole test set. This reuses the ground-truth anomaly masks for parameter selection and can lead to optimistic results given the small test sets. Appendix E.4 states that SimpleNet, SuperSimpleNet, and GLASS use default parameters without tuning. SuperSimpleNet obtains 0.3786 AUROC on CHAOS. PatchCore, EfficientAD, and DDAD have only one or two tuned parameters. These choices make it difficult to assess whether the reported improvements reflect better generalization or more favorable parameter selection.

  2. The experiments do not clearly establish the contribution of the proposed typicality score. Appendix G studies sample size, superpixel size, and feature weighting. It does not compare the typicality score in Eq. (5) with negative log-likelihood computed from the same estimated distributions. In Table 8 LACToS with M=1 already exceeds DDAD on FFHQ and DVXray although Appendix G states that it degenerates toward a reconstruction-error-like score. It is therefore unclear whether typicality itself provides an advantage over simpler distribution-aware scores or whether the gains mainly come from posterior sampling, feature augmentation, superpixels, and spatial smoothing.

3 Incorrect author information in the references. Line 548 incorrectly includes John Kieffer as the book's third author. The reference at lines 563 to 565 incorrectly includes the editor Peter Dayan as the second author.

Questions for authors
  1. Could the authors repeat the main comparisons using disjoint validation and test sets with comparable tuning opportunities for all methods? Do the conclusions still hold particularly on DVXray where PatchCore has the higher AUROC and nearly equal AP and AUPRO?

  2. How does the proposed typicality score compare with negative log-likelihood alone on the Multimodality Toy Experiment and the three main datasets with posterior samples, superpixels, features, histogram estimation, and spatial smoothing held fixed?

Reviewer_8tqoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates local anomaly detection under complex multimodal distributions. The authors propose LACToS, which leverages conditional distributions generated by diffusion models and superpixel-level statistical typicality to assess anomalies in local image regions. Experiments are conducted on medical CT scans, human face images, and X-ray security screening images, with comparisons against several anomaly detection methods. The results demonstrate that LACToS achieves competitive anomaly localization performance across multiple datasets.

Critical strengths

The paper addresses an important practical challenge in anomaly detection, which is detecting anomalies when the inlier data distribution is multimodal. The experiments also cover diverse imaging domains, providing an evaluation of the proposed method under more complex and heterogeneous anomaly detection scenarios.

Critical weaknesses
  1. The theoretical foundation of superpixel-level typicality relies on the assumption that pixels within the same superpixel are conditionally IID given the surrounding image context. However, this assumption is unlikely to hold in natural images, where neighboring pixels exhibit strong spatial correlations. Generally, superpixels are constructed based on local color or visual similarity, which also implies statistical dependence among their constituent pixels. When strong spatial correlations exist, increasing the number of pixels within a superpixel does not necessarily provide a corresponding increase in the number of independent observations, potentially undermining the statistical justification for the proposed typicality measure.

  2. LACToS performs substantially worse than existing methods on MVTec AD, raising concerns about its effectiveness as a general-purpose local anomaly detection approach. This limitation is particularly important because, in practical visual anomaly detection scenarios, the characteristics of the inlier distribution, such as whether it is unimodal or multimodal, are not known in advance. Therefore, strong and consistent performance under both unimodal and multimodal distributions is important for a broadly applicable anomaly detection method. The current experimental results do not sufficiently demonstrate such generalizability.

Questions for authors

Is there any reason or specific design in LACToS leading to the worse performance on MVTec AD?

Reviewer_9iwWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LACToS for local anomaly detection using conditional typicality rather than reconstruction error alone. It groups pixels into superpixels and uses shared diffusion denoising samples to approximate local inpainting distributions. Experiments include controlled toy settings and three applications: synthetic liver lesions, facial anomalies, and X-ray baggage screening. The reported results suggest benefits when normal image content is multimodal.

Critical strengths
  1. Well-motivated distribution-aware scoring. Conditional typicality addresses a meaningful limitation of reconstruction-error methods under multimodal normal distributions. Superpixel aggregation and shared posterior sampling make this approach computationally tractable.
  2. Promising evidence beyond industrial inspection. In Table 1, LACToS achieves the highest AP and AUPRO on all three primary benchmarks, with particularly notable gains on CHAOS AP and FFHQ AP/AUPRO. The toy experiments and sample-count ablation provide additional support for the motivation.
Critical weaknesses
  1. The tuning and test data are not clearly separated. Appendix E.3 states that hyperparameters are selected using the highest AUROC on 20 test images, and the chosen configuration is then evaluated on the whole test set. This suggests that the same images are used both to choose the parameters and to report performance, which could make the results look better than they would on unseen data. Appendix E.4 also states that SimpleNet, SuperSimpleNet, and GLASS use default parameters, while other methods are tuned for each dataset. This makes it difficult to tell how much of the improvement comes from the method itself rather than differences in tuning. The authors should clarify the data splits and report results on a separate test set, with comparable tuning efforts for the baselines.
  2. The paper gives different descriptions of the anomaly score. Equation (5) subtracts entropy from the average negative log-likelihood, but Equations (14)–(15) additionally multiply the entropy term by the number of pixels in the superpixel. There is also a difference in how RGB values and features are combined: Sections 4.3–4.4 describe building one joint histogram, whereas Appendix E.3 computes two separate typicality scores and adds them with a weight. These are not generally the same score. A consistent formula and clear pseudocode would help resolve this issue. In addition, Theorem D.1 relies on assumed bounds for the difference between denoising and inpainting distributions, and the theoretical analysis does not cover the feature component. The authors should explain which parts of the implemented method are supported by the theory.
  3. The results support a narrower robustness claim than the paper makes. Table 8 includes useful comparisons with the same number of samples, but LACToS does not improve every metric. For example, with five samples, DDAD achieves a higher AUPRO on CHAOS than LACToS (0.8827 versus 0.8764). On DVXray, the gains over PatchCore are small, and no confidence intervals or repeated-run results are provided to show whether they are reliable. The benchmarks also have limitations: the liver lesions are synthetic, normal and anomalous face images come from different sources, and the X-ray experiment only considers scissors. These experiments are useful, but they do not yet show that the method is broadly more robust across real-world anomaly types. The claims should be limited to the settings supported by the results.
Questions for authors
  1. Are the 20 images used for parameter selection also included in the final test results? Please give the size of each data split and, if they overlap, report results on a test set that was not used for tuning. How were the tuning efforts made comparable across methods?
  2. Which scoring formula and RGB–feature combination were used in the experiments? Please provide consistent equations and pseudocode, and clarify whether any correction would change the reported results.
  3. Can the benefit of typicality scoring be tested by keeping the posterior samples, features, superpixels, and postprocessing the same, and changing only the score? Can a small controlled experiment compare the shared denoising samples with samples obtained by inpainting each region separately?
  4. Were CHAOS training and test data split by patient? How were the facial anomaly masks checked, possible image-source effects assessed, and DVXray images selected or excluded? It is recommended to also provide confidence intervals or results from repeated runs for the main comparison.

Comments on the paper

1 comment

We would like to thank the reviewers for their constructive feedback. While we have decided to withdraw our paper and revise it, we respond to the main points raised. In particular:

  • Thanks for pointing out that there was overlap between the samples used for hyperparameter tuning and for method evaluation. We have corrected this by removing the tuning samples from the test set and recomputing all metrics. Our main results are all slightly perturbed after this correction due to the slightly smaller test sets, but there are no qualitative changes to their interpretation.
  • Several reviewers also pointed out the lower performance of LACToS on MVTec. This was because, at initial submission, we were using a single set of hyperparameters across all MVTec categories, whereas some baselines (e.g. DDAD) tune hyperparameters separately on each category. We have since tuned LACToS separately on each category and its performance is now comparable to (though not always better than) the baselines.
  • Regarding hyperparameter tuning, LACToS, PatchCore, EfficientAD, and DDAD were tuned on the same validation images using the same selection criterion (highest pixel-level AUROC) and similar parameter grid search granularity (though different methods have different hyperparameters). SimpleNet, SuperSimpleNet, and GLASS used their published default configurations, since tuning their training hyperparameters would require retraining. We will state this distinction and the full hyperparameter search spaces explicitly in the revised paper.
  • We have corrected inconsistencies in our description of the anomaly score. In particular, while auditing our code and previous manuscript draft, we found that Eq. (5) shows the absolute deviation from entropy (unsigned typicality), whereas our implementation used the signed deviation (signed typicality). All reported results use the signed typicality score, and we have revised the text to match. Likewise, we have reworded our claims to more precisely match the evidence in our numerical results.
  • CHAOS training and evaluation data are indeed split by patient, so that images from the same patient do not appear in both sets.
  • Regarding alternative statistical metrics beyond typicality, we have now also evaluated negative log likelihood thresholding under our same superpixel conditional distribution estimation protocol, and it also works well.
  • Regarding the assumption of conditional independence of pixels within each superpixel, we realize this is not generally true; our original submission appendix includes an analysis of our typicality score without this assumption (page 22, line 1172).

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

Related papers

Discussion (0)

Sign in to comment.