- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper combines multiscale signed-discrepancy responses with a geometry-conditioned Gaussian for 3D anomaly detection, achieving strong localization results on Real3D-AD and Anomaly-ShapeNet.
Critical strengths
The method addresses spatial patterns beyond discrepancy magnitude through a compact formulation. Strong P-AUROC results, ablations, and matched-pair analysis support its effectiveness.
Critical weaknesses
-
Novelty: What advantages does the decomposition offer over smoothing responses, DoG, or heat-kernel representations under the same Gaussian model?
-
Geometry conditioning: Can two descriptors distinguish regions with similar geometry but different normal-variation distributions?
-
Discrepancy representation: How robust is signed normal displacement to missing geometry, tangential deformations, registration errors, and partial observations?
-
Spatial dependencies: Since covariance captures only within-location scale relationships, can the method detect abnormal relationships between distant regions?
-
Validation: Does the multiscale advantage persist against gradient or Laplacian baselines, and are the 221 matched pairs representative?
-
Generalization: Can the unified model handle unseen categories using one normal reference without retraining?
Questions for authors
Please refer to the weakness
6 comments
Response to Reviewer vCN8 (2/5): W2, W3
W2. Geometry conditioning
In most cases, yes. For regions with similar geometry whose normal variation differs by at least a factor of two, the covariance predicted from the two descriptors fits a region’s own normal responses better than its partner’s in about three out of four comparisons, compared with 50% without conditioning. The limit is that locations with identical descriptors receive the same predicted distribution.
Test. We take pairs of regions whose mean descriptors are similar (, the same threshold as in Appendix B.3). For a pair of regions A and B, we evaluate the held-out normal responses of A once with the covariance predicted for A and once with the covariance predicted for B, and we count which one gives the lower NLL. Without geometry conditioning, the two covariances are identical, so the rate is 50% by construction.
Table B. Geometry conditioning. Entries report the preference rate for each region’s own predicted covariance over its partner’s, based on held-out normal NLL with predicted means fixed. Ties count as 0.5, and categories are weighted equally. (V: normal variation ; H: normal-height ratio .)
| Conditioning | Normal-region comparisons | Real3D-AD (%) | Anomaly-ShapeNet (%) |
|---|---|---|---|
| None | All matched pairs | 50 | 50 |
| V only | All matched pairs | 80.86 | 72.36 |
| H only | All matched pairs | 77.80 | 69.78 |
| V + H | All matched pairs | 81.75 | 76.86 |
| V + H | Estimated variation ratio ≥ 2 | 77.09 | 74.23 |
All matched pairs. With both descriptors, the region’s own covariance gives the lower NLL in 81.75% of the comparisons on Real3D-AD and in 76.86% on Anomaly-ShapeNet. Using both descriptors raises this rate by 0.89 pp and 4.50 pp, respectively, over the best single descriptor.
Similar geometry, different normal variation. This is the case raised in the question. We restricted the pairs to those whose estimated normal variation differs by at least a factor of two, where the variation is the trace of the response covariance across three normal scans, averaged within each region. The region’s own covariance still gives the lower NLL in 77.09% of the comparisons on Real3D-AD (239 pairs, 12 categories) and in 74.23% on Anomaly-ShapeNet (332 pairs, 35 categories).
W3. Discrepancy representation.
Signed normal displacement measures geometric deviations at observed locations. We address the four cases separately below.
Missing geometry. In the default setting, scores are evaluated only at observed query points. A nominal region with no query points is therefore not scored directly, which makes the default setting less sensitive to missing geometry (Hole P-AUROC: 63.74%). To check whether this comes from where the scores are read out or from the representation itself, we kept the trained model fixed and changed only the scoring step. There, we also used the information on which nominal regions are not covered by the query, and assigned it to the observed points within four mesh edges (Table C).
Table C. Scoring with nominal-surface coverage.
| Scoring setting | Anomaly-ShapeNet (hole defects) O / P (%) | Anomaly-ShapeNet (all) O / P (%) |
|---|---|---|
| Default (observed points only) | 58.42 / 63.74 | 94.75 / 97.20 |
| With nominal-surface coverage | 61.75 / 79.39 | 94.69 / 97.86 |
With coverage included, Hole P-AUROC rises from 63.74% to 79.39%, while the overall O/P-AUROC stays at the same level (−0.06 / +0.66 pp). Since the trained model is unchanged, the lower sensitivity to missing geometry comes in large part from the default readout, and not from the representation itself. All results in the paper use the default setting.
Response to Reviewer vCN8 (3/5): W3 (continued)
Tangential deformations. A tangential feature shift can change the surface geometry and thus produce a normal discrepancy. We tested this by shifting a hole feature along its supporting surface on four HPPF Industry CAD models, by and , where is the original feature width. We used the default energy-gap score, with the nominal surface and trained model fixed.
Table D. Tangential feature shifts on CAD models.
| CAD case | P-AUROC (%) | P-AUROC (%) |
|---|---|---|
| Bearing Cap | 91.52 | 93.26 |
| Dynamixel Plate | 82.88 | 91.22 |
| Franka Connector | 79.12 | 86.81 |
| Teeth Wrist Servo | 91.44 | 95.83 |
| Mean | 86.24 | 91.78 |
These controlled CAD results use distance-based proxy masks: before registration, points are labeled anomalous when their distance to the original clean CAD surface exceeds , where D is its bounding-box diagonal. O-AUROC is not reported because independent normal test scans are unavailable. Motion that leaves the surface geometry unchanged provides no geometric discrepancy to detect.
Registration errors. We tested two aspects. First, we injected a controlled rigid error after alignment, without re-alignment, and evaluated on the full test sets (Table E). For a residual rotation of or a translation of , P-AUROC decreases by less than 1 pp on both benchmarks. For larger errors, the drop is larger: at or , P-AUROC decreases by 1.5 and 2.4 pp on Real3D-AD, and by 1.4 and 3.9 pp on Anomaly-ShapeNet.
Second, we checked how often the registration fails to correct a wrong pose. We perturbed the initial pose before alignment, by up to in rotation and in translation (720 trials on Real3D-AD and 2,400 on Anomaly-ShapeNet). The registration returns to the reference pose in 97.8% of the trials on Real3D-AD and in 94.7% on Anomaly-ShapeNet. Moreover, in these trials the result typically reproduces the reference registration: the median residual rotation is below degrees and the median residual translation is below . Thus, although the score is sensitive to a residual registration error, the registration recovers from large initial pose errors in most cases.
Table E. Residual registration error after alignment (all categories, full test sets; O / P-AUROC, %). Throughout these perturbation experiments, D denotes the bounding-box diagonal of the unperturbed nominal mesh. ΔP is computed from the displayed P-AUROC values.
| Condition | Real3D-AD O / P | ΔP | Anomaly-ShapeNet O / P | ΔP |
|---|---|---|---|---|
| Clean | 84.12 / 92.71 | – | 94.75 / 97.20 | – |
| Rotation 0.25° | 84.16 / 92.62 | −0.09 | 94.72 / 97.15 | −0.05 |
| Rotation 0.5° | 84.07 / 92.33 | −0.38 | 94.33 / 96.90 | −0.30 |
| Rotation 1° | 83.53 / 91.19 | −1.52 | 93.53 / 95.77 | −1.43 |
| Translation 0.1% D | 84.07 / 92.61 | −0.10 | 94.66 / 97.07 | −0.13 |
| Translation 0.25% D | 83.84 / 92.07 | −0.64 | 94.45 / 96.23 | −0.97 |
| Translation 0.5% D | 83.10 / 90.32 | −2.39 | 93.45 / 93.23 | −3.97 |
Partial observations. The observation mask keeps the measured residuals fixed, and the harmonic extension fills the unobserved regions for the surface solves. To test this case, we reduced the surface coverage of each query to 90% and 75%, and, for comparison, the point density to the same levels. In each condition, we reran the pipeline from registration onward (Table F). Coverage is measured relative to the surface area observed in the original query, not to the entire nominal surface. If an anomalous scan loses all of its defect-labeled points after cropping, we exclude it from O-AUROC.
Table F. Density and coverage (6 Real3D-AD categories: Airplane, Candybar, Fish, Gemstone, Shell, Starfish; 15 Anomaly-ShapeNet categories; O / P-AUROC, %).
| Condition | Real3D-AD (6) | Anomaly-ShapeNet (15) |
|---|---|---|
| Clean | 84.67 / 89.84 | 94.67 / 97.51 |
| Density 90% | 88.67 / 89.20 | 94.40 / 97.56 |
| Density 75% | 86.89 / 89.56 | 94.04 / 97.55 |
| Coverage 90% | 83.13 / 89.92 | 93.34 / 97.19 |
| Coverage 75% | 83.11 / 89.44 | 87.11 / 95.72 |
Under lower density, P-AUROC changes by at most 0.64 pp on Real3D-AD and by less than 0.1 pp on Anomaly-ShapeNet. Under reduced coverage, the Real3D-AD results change little (−1.6 pp O-AUROC and −0.4 pp P-AUROC at 75%). On Anomaly-ShapeNet, 75% coverage reduces O-AUROC by 7.6 pp and P-AUROC by 1.8 pp. Thus, localization is largely preserved under partial observations, while object-level detection on Anomaly-ShapeNet is more sensitive to reduced coverage.
Response to Reviewer vCN8 (1/5): W1
We thank the reviewer for the thoughtful questions, and for recognizing the compact formulation and the supporting analyses. Following these questions, we ran several additional comparisons and diagnostics. We address each weakness below.
Following your suggestion, we compared two choices separately: (1) whether to use the smoothed responses or their differences, and (2) which smoothing operator to use. In all four-channel variants, we kept the field lifting, the geometry descriptors, and the scorer architecture fixed, and we trained each scorer separately under the same normal-only protocol.
Table A. Response representations under the same scoring protocol (O / P-AUROC, %). All multiscale variants use four channels; residual-only scoring uses one. Scales are ℓj=2jh0 for j=0,1,2. Heat filters use tj=(ln2)ℓj2. Geodesic Gaussian filters use σj=2tj with a 3σj cutoff, and spectral-wavelet peaks are matched to λj=1/ℓj2. SP: screened-Poisson.
(1) Differences versus smoothed responses. The screened-Poisson (SP) smoothing stack [f,u0,u1,u2] and our bands contain the same information, since each can be recovered from the other. The difference lies in how the normal model is fitted. We fit a separate model in each representation, and the standardization, the covariance shrinkage, and the variance floor are applied in the coordinates of that representation. In the band form, each change between two scales is itself a coordinate, so it is standardized and regularized directly. In our experiments, this leads to higher accuracy: the bands improve O/P-AUROC over the stack by 9.9 / 3.8 pp on Real3D-AD and by 0.5 / 0.6 pp on Anomaly-ShapeNet. We therefore see the bands as better coordinates for fitting the normal model, and not as a representation with more information.
(2) Screened-Poisson versus other kernels. In our experiments, the SP bands are the most accurate on both metrics and both benchmarks. Compared with the best of the three alternative kernels for each metric, they are higher by 2.7 / 2.6 pp (O / P) on Real3D-AD and by 1.1 / 0.3 pp on Anomaly-ShapeNet. To be clear, we do not claim that other kernels cannot work. All three alternative kernels improve P-AUROC over residual-only scoring, showing that the localization benefit extends beyond a single operator.