- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper studies the problem where one normal mode may have only a few reference samples. Normal samples from this data-rare mode then receive higher anomaly scores under the global covariance metric. More of them exceed the threshold calibrated at full coverage and lead to increased false positives. The authors propose a two-mode mixture model to separate within-mode covariance shifts from between-mode mean differences. They decompose this score increase along the eigenvectors of the full-coverage covariance. Most of the increase comes from the leading eigenvectors in DINOv3 ViT-B/16. On MVTec AD they compare targeted deletion with matched random deletion and reweighting to distinguish composition from sample count. They also design Head-Calibrated Mahalanobis (HCM) based on these findings to reduce this false-positive inflation at the cost of lower detection performance.
Critical strengths
The problem studied in this paper is valuable for anomaly detection. Normal reference sets can contain several valid modes such as different poses or illumination conditions. These modes can have very different numbers of samples. The resulting false positives on data-rare modes can be hidden by aggregate detection metrics such as image AUROC. The authors split the score increase behind these false positives into exact contributions from each covariance eigenvector. These contributions sum to the total increase and show which eigenvectors drive it.
Critical weaknesses
W1. Inconsistent reporting of the main results. At q = 0.5 the HCM I-AUROC is 0.9061 in Table 10 but 0.925 in Table 11. Table 11 only recalibrates the threshold while I-AUROC depends solely on the ranking of image scores. Mahalanobis has the same I-AUROC in both tables (0.9514 vs 0.951) so the two HCM rows differ in their scores and not only in the threshold. This undermines the claim in Section 6 that HCM helps mainly at fixed thresholds. I am also confused by Appendix A because a holdout reanalysis gives different estimates from the selected primary results. It is unclear which set supports the claims and the primary gains lack uncertainty estimates.
W2. Dependence of the leading-rank taper on the residual scale. Eqs. (9)–(10) imply that multiplying all residuals by scales the contribution of eigenvector by . With the taper this factor varies across eigenvectors and can change the ranking of image scores. Without the taper () it is a common factor. The ratio stated after Eq. (10) exceeds one only when . The residual scale thus decides whether the taper raises or lowers the leading-rank weights. I could not find how in Section 3 is constructed or normalized. This weakens the validity of the HCM results.
Questions for authors
Q1. Do Tables 10 and 11 use the same HCM configuration and the same image scores? I suggest that the authors report a matched comparison in which the fixed and recalibrated thresholds are applied to the same scores.
Q2. How are the residuals in Section 3 constructed and normalized? I suggest that the authors rescale the same residuals by several factors and check whether the reduction in target persists. For each the threshold can be calibrated at and reused at reduced coverage.
Q3. Table 4 reports only the differences between targeted deletion and each control. If its targeted values match Table 3 then matched random deletion gives a target of 0.179/0.374/0.496 at the three reduced coverages. This is 67%–79% of the targeted effect while fixed-support reweighting gives only 0.016/0.058/0.155. I suggest that the authors report the absolute target of each control.
Q4. Does the leading 15% band in Table 1 cover the same ranks as the first-15% HCM head in Table 2? Both tables report fixed-residual attribution on DINOv3 ViT-B/16 as category–target–seed means. Yet the head contribution is 123.28 to 245.30 in Table 1 and 5550.17 in Table 2.