acceptodds
Under review as a conference paper at ICLR 2027

Multiscale Surface Equilibrium Responses for Unified 3D Anomaly Detection

Abstract

Two geometric deviations of the same magnitude can carry different evidence of a defect: a narrow bump is often anomalous where a broad variation is not. Reliable 3D anomaly detection therefore requires interpreting not only the size of a deviation but its spatial pattern, relative to the variation observed in normal shapes. We propose a unified framework that uses a nominal surface as the analysis domain and represents each query through four multiscale equilibrium responses. Signed query-to-surface discrepancies are lifted into a scalar field, and screened- Poisson equilibria reveal how this field responds to increasing spatial coupling. This four-channel decomposition is exactly additive: it preserves the input discrepancy while exposing its structure across spatial scales. A single geometry- conditioned Gaussian potential, shared across categories and trained only on nor mal scans, evaluates the joint responses against the variation expected at each surface location. Anomaly scores measure the energy gap from the expected nor mal response, assigning high cost to departures along directions with little normal variation. Despite using one shared model per dataset, our method surpasses not only unified baselines but also the strongest category-specific methods: it achieves point-level AUROC of 92.7% and 97.2% on Real3D-AD and Anomaly-ShapeNet, exceeding the previous best by 4.9 and 3.6 points, respectively, while also attaining the best mean object-level among unified methods on both benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Anomaly Detection, Computer Vision, 3D

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_vCN8OpenReview ↗
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
  1. Novelty: What advantages does the decomposition offer over smoothing responses, DoG, or heat-kernel representations under the same Gaussian model?

  2. Geometry conditioning: Can two descriptors distinguish regions with similar geometry but different normal-variation distributions?

  3. Discrepancy representation: How robust is signed normal displacement to missing geometry, tangential deformations, registration errors, and partial observations?

  4. Spatial dependencies: Since covariance captures only within-location scale relationships, can the method detect abnormal relationships between distant regions?

  5. Validation: Does the multiscale advantage persist against gradient or Laplacian baselines, and are the 221 matched pairs representative?

  6. Generalization: Can the unified model handle unseen categories using one normal reference without retraining?

Questions for authors

Please refer to the weakness

6 comments
AuthorsOpenReview ↗

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.

W1. Advantage over other representations.

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\ell_j=2^j h_0 for j=0,1,2j=0,1,2. Heat filters use tj=(ln⁡2)ℓj2t_j=(\ln 2)\ell_j^2. Geodesic Gaussian filters use σj=2tj\sigma_j=\sqrt{2t_j} with a 3σj3\sigma_j cutoff, and spectral-wavelet peaks are matched to λj=1/ℓj2\lambda_j=1/\ell_j^2. SP: screened-Poisson.

RepresentationReal3D-ADAnomaly-ShapeNet
Residual only [f][f]72.3 / 80.293.7 / 93.6
SP smoothing stack [f,u0,u1,u2][f, u_0, u_1, u_2]74.2 / 88.994.3 / 96.6
Heat-kernel differences81.4 / 90.193.5 / 96.9
Geodesic Gaussian DoG81.2 / 90.193.6 / 96.8
Spectral graph wavelets76.9 / 87.393.7 / 96.7
SP bands (Ours)84.1 / 92.794.8 / 97.2

(1) Differences versus smoothed responses. The screened-Poisson (SP) smoothing stack [f,u0,u1,u2][f, u_0, u_1, u_2] 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.

AuthorsOpenReview ↗

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 (hnv,hnh)(h_{nv}, h_{nh}) are similar (dh≤0.5d_h \le 0.5, 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 hnvh_{nv}; H: normal-height ratio hnhh_{nh}.)

ConditioningNormal-region comparisonsReal3D-AD (%)Anomaly-ShapeNet (%)
NoneAll matched pairs5050
V onlyAll matched pairs80.8672.36
H onlyAll matched pairs77.8069.78
V + HAll matched pairs81.7576.86
V + HEstimated variation ratio ≥ 277.0974.23

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 settingAnomaly-ShapeNet (hole defects) O / P (%)Anomaly-ShapeNet (all) O / P (%)
Default (observed points only)58.42 / 63.7494.75 / 97.20
With nominal-surface coverage61.75 / 79.3994.69 / 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.

AuthorsOpenReview ↗

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 0.10w0.10w and 0.25w0.25w, where ww 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 case0.10w0.10w P-AUROC (%)0.25w0.25w P-AUROC (%)
Bearing Cap91.5293.26
Dynamixel Plate82.8891.22
Franka Connector79.1286.81
Teeth Wrist Servo91.4495.83
Mean86.2491.78

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 10−5D10^{-5}D, 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 0.5∘0.5^\circ or a translation of 0.25% D0.25\%\,D, P-AUROC decreases by less than 1 pp on both benchmarks. For larger errors, the drop is larger: at 1∘1^\circ or 0.5% D0.5\%\,D, 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 90∘90^\circ in rotation and 10% D10\%\,D 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 10−310^{-3} degrees and the median residual translation is below 10−7% D10^{-7}\%\,D. 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.

ConditionReal3D-AD O / PΔPAnomaly-ShapeNet O / PΔP
Clean84.12 / 92.71–94.75 / 97.20–
Rotation 0.25°84.16 / 92.62−0.0994.72 / 97.15−0.05
Rotation 0.5°84.07 / 92.33−0.3894.33 / 96.90−0.30
Rotation 1°83.53 / 91.19−1.5293.53 / 95.77−1.43
Translation 0.1% D84.07 / 92.61−0.1094.66 / 97.07−0.13
Translation 0.25% D83.84 / 92.07−0.6494.45 / 96.23−0.97
Translation 0.5% D83.10 / 90.32−2.3993.45 / 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, %).

ConditionReal3D-AD (6)Anomaly-ShapeNet (15)
Clean84.67 / 89.8494.67 / 97.51
Density 90%88.67 / 89.2094.40 / 97.56
Density 75%86.89 / 89.5694.04 / 97.55
Coverage 90%83.13 / 89.9293.34 / 97.19
Coverage 75%83.11 / 89.4487.11 / 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.

Reviewer_NDxAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a unified normal‑only 3D anomaly‑detection framework analyzing explicit query‑to‑reference surface discrepancies. Per‑category nominal meshes are built from normal scans; aligned query residuals are lifted to a scalar field, and screened‑Poisson equations at multiple scales produce four exactly‑additive response channels. A dataset‑shared geometry‑conditioned Gaussian model predicts channel means and full covariances from local surface descriptors. Point anomaly scores are energy gaps relative to expected normal responses, and object scores average the top‑1% point scores. It establishes discrete‑solution uniqueness, non‑amplifying spectral behaviour and exact discrepancy reconstruction. On Real3D‑AD and Anomaly‑ShapeNet, it attains 92.7% and 97.2% point‑level AUROC, with 84.1% and 94.8% object‑level AUROC, the best among unified methods. Ablations confirm the value of multiscale representation, geometry conditioning and full covariance; matched‑amplitude analysis verifies it separates local defects from broad normal variations beyond simple residual‑magnitude thresholding.

Critical strengths
  1. Screened-Poisson smoothing and conditional Gaussian modeling are individually established tools, but the paper derives an exactly reconstructing hierarchy of equilibrium differences on a nominal surface and uses their joint, geometry-conditioned covariance as anomaly evidence. This yields a specific representation of the paper's central object—spatially organized geometric discrepancy—and is a meaningful adaptation beyond applying an off-the-shelf filter bank.
  2. One shared model achieves state-of-the-art point localization on two different benchmarks, while multiscale, conditioning, covariance, single-band, and matched-amplitude experiments test whether the benefit comes from spatial scale and conditional normal variation rather than raw residual magnitude.
Critical weaknesses
  1. The Shell failure case, where repeated ridges and alignment/sampling offsets cause severe false positives and object AUROC falls to 37.1%, is not peripheral: residuals enter the anomaly score directly. Without controlled registration and surface-reconstruction perturbations, it is unclear how robust the reported gains are under realistic pose, coverage, and scanning variation.
  2. The paper evaluates only two established benchmarks and does not clearly state whether the scale count, interaction lengths, neighborhood size, covariance choices, and top-1% object aggregation were selected exclusively using normal validation data. Cross-dataset fixed-hyperparameter evaluation, runtime measurements, and reference-selection sensitivity would materially strengthen the generalization claim.
Questions for authors
  1. How were the number of scales, interaction lengths, geometry‑neighborhood size, covariance regularization, and top‑1% object aggregation selected? Were any anomaly labels from either benchmark used directly or indirectly? Could the authors elaborate on the exact normal‑only validation protocol, and show cross‑dataset results under fixed hyper‑parameters?
  2. How does performance change under controlled initial rotations/translations, ICP failure, nominal‑mesh reconstruction error, point‑density changes, partial coverage, and alternative choices of the single nominal reference? I would like to see performance‑versus‑perturbation curves, alongside the fraction of queries for which registration is deemed unreliable.
  3. What are the end‑to‑end preprocessing, training, and per‑query runtimes, including NKSR reconstruction, ICP, lifting, screened‑Poisson solves, and scoring? Could the authors also specify the full ICP and NKSR settings needed for reproduction?
  4. How much of the gain is specific to the screened‑Poisson construction rather than to any multiscale smoothing bank? Could the authors include comparisons against compute‑matched Gaussian/geodesic smoothing, Laplacian‑pyramid or spectral‑wavelet responses, all fed into the identical conditional full‑covariance scorer.
  5. The conditional response distribution is modeled as a single Gaussian. How well calibrated is this assumption across categories and surface regions? I would expect held‑out normal likelihood and calibration diagnostics, alongside comparisons against mixture‑model, robust‑covariance, or non‑parametric conditional‑density alternatives under matched parameter budgets.
  6. How stable are results across the choice of nominal reference scan and random seed? Could the authors report mean and variance aggregated over multiple reference selections, with particular attention to categories featuring repeated geometry or substantial intrinsic normal‑shape variation?
9 comments
AuthorsOpenReview ↗

Response to Reviewer NDxA (1/5): Q1, W2

We thank the reviewer for the careful and detailed review, and for recognizing that our multiscale representation goes beyond applying an off-the-shelf filter bank. The requested experiments helped us to state the scope and the limitations of the method more precisely. We address each point below.

Q1, W2. Hyperparameter selection and validation

Most values were fixed in advance rather than tuned, and the only data-driven choices used held-out normal scans. No anomaly labels from either benchmark were used for training, for epoch selection, or for choosing the hyperparameters below. We used one common configuration on both benchmarks. The test-AUROC sweeps reported below assess sensitivity and were not used for hyperparameter selection.

Model selection. Each category has four normal scans, one of which is the reference; the three folds each hold out one of the other three. The reference scan stays in training, and the normalization statistics are computed from the training normals of that fold only. We select the epoch with the lowest mean held-out normal NLL across folds, and then train a new model on all normal training scans for that number of epochs.

Scales. The interaction lengths are not tuned. They double between successive scales, ℓj=2jh0\ell_j = 2^j h_0, and h0=max⁡(2hmesh,hsampling)h_0=\max(2h_{\mathrm{mesh}},h_{\mathrm{sampling}}) is computed from the mesh spacing and sampling resolution of the normal reference. We set J=3J=3 to obtain a compact four-channel representation: a fine-scale residual, two intermediate differences, and the coarsest field. With J=2J=2, only one intermediate difference remains, whereas a larger JJ increases the response dimension and the covariance size. We do not claim that J=3J=3 is optimal. The sensitivity analysis in Table 5b shows that the results are stable: for J=2J=2 to 55, P-AUROC varies by at most 0.7 pp on Real3D-AD and 0.4 pp on Anomaly-ShapeNet.

Neighborhood size. We selected K=512K = 512 by held-out normal NLL, which requires no anomaly labels (Table A). Among the tested values K∈{128,256,512,1024}K \in \lbrace 128, 256, 512, 1024 \rbrace, it gives the lowest NLL on both benchmarks: −25.81 on Real3D-AD and −29.27 on Anomaly-ShapeNet. We therefore use the same neighborhood size on both benchmarks.

Object aggregation. Top-1% averaging was fixed from the outset, in order to keep localized evidence without relying on a single extreme point. Its O-AUROC is within 0.80 pp of the best tested aggregation on Real3D-AD and within 0.10 pp on Anomaly-ShapeNet (Table B). All aggregations share the same point scores, so P-AUROC is unchanged.

Covariance regularization. We fixed the diagonal shrinkage at α=0.1\alpha=0.1 and the additive variance floor at 0.052I0.05^2 I to stabilize fitting in the standardized response coordinates. Performance changes little across the three tested settings (Table C): O/P-AUROC varies by only 0.23/0.09 pp on Real3D-AD and 0.05/0.05 pp on Anomaly-ShapeNet.

Cross-dataset evaluation. Finally, as requested, we tested cross-dataset transfer (Table D). This test is stricter than fixing the hyperparameters only: we also kept the source-model weights and normalization statistics, without retraining or adaptation on the target data; only the category-specific nominal surfaces come from the target normal references. Real3D-AD → Anomaly-ShapeNet reaches 94.43 / 97.08 (O / P), which is within 0.32 / 0.12 pp of the in-dataset model. Anomaly-ShapeNet → Real3D-AD reaches 74.36 / 90.69. Here localization is largely preserved (−2.02 pp), but object-level detection drops by 9.76 pp.

Table A. Neighborhood-size sensitivity measured by held-out normal NLL (nats).

KReal3D-AD NLL ↓Anomaly-ShapeNet NLL ↓
128−25.5795−29.1614
256−25.6237−29.1982
512 (default)−25.8095−29.2728
1024−25.7497−29.2542

Table B. Object-level aggregation sensitivity.

AggregationReal3D-AD O-AUROC (%) ↑Anomaly-ShapeNet O-AUROC (%) ↑
Max84.90994.446
Top 0.5%84.24694.848
Top 1% (default)84.11694.753
Top 5%82.11994.079
Mean79.11292.017

Table C. Covariance regularization sensitivity. The additive variance floor is ϵ2I\epsilon^2 I. Results are O / P-AUROC (%).

α\alphaϵ\epsilonReal3D-AD O / P ↑Anomaly-ShapeNet O / P ↑
00.0583.93 / 92.6394.70 / 97.15
0.10.0584.12 / 92.7194.75 / 97.20
0.10.02584.16 / 92.7294.74 / 97.20

Table D. Cross-dataset generalization. Differences are measured against our model trained on the target dataset.

Source → TargetO-AUROC (%)P-AUROC (%)Δ vs. target-trained model (O / P, pp)
AnomalyShapeNet → Real3D-AD74.3690.69−9.76 / −2.02
Real3D-AD → AnomalyShapeNet94.4397.08−0.32 / −0.12
AuthorsOpenReview ↗

Response to Reviewer NDxA (2/5): W1, Q2 (1)

W1. Shell failure case

We agree that the Shell failure is not a peripheral case, and that registration and nominal-surface quality affect the residuals used by our score. We therefore quantify these dependencies through the controlled perturbation experiments in Q2.

In Shell, the median object score for anomalous scans is only 0.63 times that for normal scans, compared with 14.04 when scans from the other eleven categories are pooled (Table E). This reversal is consistent with Shell’s below-chance O-AUROC of 37.1%. Changing the nominal reference does not resolve the failure: Shell’s mean O-AUROC remains 36.9%, with a standard deviation of 0.62 pp (Q6).

A plausible explanation for the high scores on normal scans is that the repeated ridges make the registration ambiguous, so that slightly shifted alignments produce apparent discrepancies even on normal surfaces. The initial-pose experiment in Q2 is consistent with this explanation. On Real3D-AD, 15 of the 16 unrecovered trials occur in Shell: it recovers in 45 of 60 trials, while the other eleven categories recover in 659 of 660 trials (99.8%). Together, these results show that Shell exhibits reversed object-score ordering and accounts for nearly all registration-recovery failures on Real3D-AD under the tested initial perturbations. We have expanded Appendix C.5 to include these findings.

Table E. Object-score ordering under the default configuration. The ratio is the median anomalous object score divided by the median normal object score. Scores from the remaining eleven categories are pooled; their O-AUROC is category-macro averaged.

GroupNormal / anomalous scansAnomalous-to-normal median-score ratioO-AUROC (%)
Shell52 / 480.6337.1
Remaining 11 categories552 / 55414.0488.4 (mean)

Q2. Robustness to perturbations

We ran controlled perturbations for each factor that you listed, and we report them in five parts: (1) residual registration errors, (2) recovery from perturbed initial poses, (3) point density and coverage, (4) the nominal surface, and (5) the choice of the nominal reference. In short, small residual registration errors change P-AUROC only slightly, large initial pose errors are corrected in most trials, and point-level performance changes little under lower density and across reference choices. The method is more sensitive to larger registration errors, to perturbations of the nominal surface, and to reduced coverage on Anomaly-ShapeNet, and the Shell category remains a limitation. In experiments (1)–(4), the model weights, normalization statistics, and hyperparameters are kept fixed. Since each perturbation condition requires separate processing and evaluation, for experiments (3) and (4) we report the categories completed so far, as specified in each table.

(1) Sensitivity to residual registration errors. We inject a controlled rigid error after alignment, without re-alignment, and evaluate on the full test sets (Table F). Rotation and translation errors are injected separately.

Table F. 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 entire nominal reference. ΔP is computed from the displayed P-AUROC values.

ConditionReal3D-AD O / PΔPAnomaly-ShapeNet O / PΔP
Clean84.12 / 92.71–94.75 / 97.20–
Rotation 0.25°84.16 / 92.62−0.0994.72 / 97.15−0.05
Rotation 0.5°84.07 / 92.33−0.3894.33 / 96.90−0.30
Rotation 1°83.53 / 91.19−1.5293.53 / 95.77−1.43
Translation 0.1% D84.07 / 92.61−0.1094.66 / 97.07−0.13
Translation 0.25% D83.84 / 92.07−0.6494.45 / 96.23−0.97
Translation 0.5% D83.10 / 90.32−2.3993.45 / 93.23−3.97

For a rotation of 0.5° or a translation of 0.25% D, P-AUROC decreases by less than 1 pp on both benchmarks. For larger errors, the drop is clearly larger: at 1° or 0.5% D, P-AUROC decreases by 1.5 and 2.4 pp on Real3D-AD, and by 1.4 and 3.9 pp on Anomaly-ShapeNet. These results confirm the dependence on registration and quantify its effect.

AuthorsOpenReview ↗

Response to Reviewer NDxA (3/5): Q2 (2)–(5)

(2) Recovery from perturbed initial poses. Next, we perturb the initial pose before alignment and check whether the registration recovers. We use five normal and five anomalous scans per category and six conditions per scan, which gives 720 trials on Real3D-AD and 2,400 on Anomaly-ShapeNet (Table G). As the reference pose, we use the registration result of the same query without any injected perturbation. After accounting for the injected transform, a trial counts as recovered when the rotation error is at most 1° and the translation error is at most 0.5% D.

Table G. Recovery from initial pose perturbations (recovered trials / all trials).

Initial perturbationReal3D-ADRateAnomaly-ShapeNetRate
Rotation 15°117 / 12097.5%379 / 40094.8%
Rotation 45°117 / 12097.5%376 / 40094.0%
Rotation 90°115 / 12095.8%371 / 40092.8%
Translation 1% D119 / 12099.2%383 / 40095.8%
Translation 5% D118 / 12098.3%381 / 40095.3%
Translation 10% D118 / 12098.3%383 / 40095.8%
All six conditions704 / 72097.8%2,273 / 2,40094.7%

Under this criterion, 97.8% of the trials recover on Real3D-AD and 94.7% on Anomaly-ShapeNet, even though the perturbations reach 90° in rotation and 10% of D in translation. Among the recovered trials, the median residual rotation is below 10−310^{-3} degrees and the median residual translation is below 10−7%10^{-7}\% of D on both benchmarks. Our pipeline does not include a per-query reliability detector, so we report these rates as our estimate of unreliable registration under controlled perturbations. On Real3D-AD, 15 of the 16 unrecovered trials occur in Shell, as discussed in W1.

(3) Point density and coverage. We reduce point density by random subsampling and surface coverage by contiguous cropping, retaining 90% or 75% of the original points or observed area, respectively. We then rerun the pipeline from registration onward (Table H). 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 H. Density and coverage (6 Real3D-AD categories: Airplane, Candybar, Fish, Gemstone, Shell, Starfish; 15 Anomaly-ShapeNet categories; O / P-AUROC, %).

ConditionReal3D-AD (6)Anomaly-ShapeNet (15)
Clean84.67 / 89.8494.67 / 97.51
Density 90%88.67 / 89.2094.40 / 97.56
Density 75%86.89 / 89.5694.04 / 97.55
Coverage 90%83.13 / 89.9293.34 / 97.19
Coverage 75%83.11 / 89.4487.11 / 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, however, 75% coverage reduces O-AUROC by 7.6 pp and P-AUROC by 1.8 pp. These results indicate larger sensitivity to reduced coverage on Anomaly-ShapeNet, particularly for object-level detection.

(4) Perturbation of the nominal surface. We perturb the nominal surface with a smooth displacement along the normal, with an area-weighted RMS of 0.05% or 0.10% of D (Table I). Within each category, we scale the same displacement field to the two target amplitudes. For each perturbed surface, we recompute the geometry descriptors and operators. This is a controlled nominal-surface perturbation, not a measurement of the actual NKSR reconstruction error.

Table I. Nominal-surface perturbation (O / P-AUROC, %). Starting from Gaussian noise at the original mesh vertices, we apply 100 smoothing steps. Each step averages the current vertex value and its mesh-edge neighbor mean with weights 0.5 and 0.5. We then remove the area-weighted mean and normalize the field to unit area-weighted RMS before scaling it to each target amplitude.

RMS / DReal3D-AD (12 / 12 categories)Anomaly-ShapeNet (15 / 40 categories)
Clean84.12 / 92.7194.67 / 97.51
0.05%85.20 / 90.5195.47 / 97.08
0.10%83.20 / 89.4895.73 / 96.72

P-AUROC decreases by 2.2 and 3.2 pp on Real3D-AD, and by 0.4 and 0.8 pp on Anomaly-ShapeNet. O-AUROC decreases by 0.9 pp on Real3D-AD at 0.10% and does not decrease in the other three conditions.

(5) Choice of the nominal reference. Finally, we use other normal scans as the nominal reference (the per-category results are given in Q6). On Real3D-AD, the standard deviation across reference choices is at most 0.62 pp in O-AUROC and 0.15 pp in P-AUROC for every category, and below 0.04 pp for the category mean.

Reviewer_ii6vOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a unified 3D anomaly detection framework that distinguishes geometric deviations by their magnitude and spatial scale. It maps query-to-reference discrepancies onto a nominal surface and decomposes them into multiscale responses using screened-Poisson operators. A geometry-conditioned Gaussian model, shared across categories and trained on normal scans, scores departures from expected responses. Experiments on Real3D-AD and Anomaly-ShapeNet report strong performance.

Critical strengths

The paper is motivated by a clear observation: geometric deviations of similar size can indicate different things depending on their spatial extent. The method shows strong anomaly localization performance.

Critical weaknesses

While the framework is well motivated, it is simply a combination of existing modules. The experiments are missing an evaluation of the capability of the method in distinguishing large scale anomalies from normal variations.

Questions for authors

Can the proposed framework detect large-scale structural anomalies? For example, HPPF [1] presents industrial cases involving displaced planar regions and gear-angle misalignments. How would the proposed method distinguish these defects from normal variations? [1] Kang et al. Hierarchical Point-Patch Fusion with Adaptive Patch Codebook for 3D Shape Anomaly Detection. CVPR 2026.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ii6v (1/2): W1

We thank the reviewer for the constructive feedback, and for recognizing the motivation of our framework and its localization performance. We also appreciate the pointer to the HPPF industrial cases. We address each point below.

W1. Contribution beyond existing modules.

We agree that the screened-Poisson operator and Gaussian density models are already well known, and we do not claim either of them as new. Our contribution lies in what these components are applied to, and in what we treat as evidence of a defect.

What is new. Most prior 3D-AD methods compare a query with a learned reconstruction or embedding of it. As a result, the spatial extent of a deviation is not represented explicitly, and the score tends to follow the deviation magnitude. In contrast, we place the measured discrepancy on a nominal surface as a scalar field. On this field, a single variational objective, solved at several spatial scales, gives a set of responses that sum exactly to the input. The evidence of a defect is then the joint pattern of these responses, compared with the normal variation expected at the same local geometry.

Why this formulation matters. The multiscale representation exposes the spatial pattern of the discrepancy, while the conditional model determines whether that pattern is unusual for the local geometry. Without the multiscale responses, the Gaussian model scores a geometry-conditioned scalar residual and has no access to its spatial organization; this variant reaches 80.2% P-AUROC on Real3D-AD (Table 5a). Conversely, without the learned normal model, each band responds to normal structures as well as to defects, as Fig. 5 illustrates. Using all bands improves over any individual band: the full model reaches 92.7%, whereas the best single band reaches 89.5% (Table 5c).

Evidence that this matters. The matched-pair analysis (Sec. 4.3) tests this point directly. In these 221 pairs, the normal and anomalous regions have almost the same magnitude (median ratio 1.047). A magnitude-only rule ranks the anomaly higher in only 60.6% of the pairs, while our score does so in 92.8%.

Additional experiment. To examine whether the gain extends beyond one particular filter, we replaced our operator with other multiscale filter banks, using the same geometry conditioning and scorer architecture, with each scorer fitted separately under the same normal-only protocol (Table A). First, every alternative improves substantially over the residual alone (87.3–90.1% vs. 80.2% P-AUROC on Real3D-AD), which suggests that the formulation itself carries a large part of the gain. Second, the screened-Poisson responses are still the most accurate (92.7%), so the specific operator contributes as well.

Table A. Comparison of multiscale response operators.

ResponseReal3D-AD P-AUROCAnomaly-ShapeNet P-AUROC
Residual only [f][f]80.293.6
Heat-kernel differences90.196.9
Geodesic Gaussian DoG90.196.8
Spectral graph wavelets87.396.7
Screened-Poisson responses (Ours)92.797.2
AuthorsOpenReview ↗

Response to Reviewer ii6v (2/2): Q1

Q1. Large-scale structural anomalies

Our method does not treat a deviation as normal simply because it is broad. The decomposition retains the full discrepancy, including its coarse-scale content, and all scale responses are scored jointly against the normal distribution predicted for the local geometry. A broad structural deviation can therefore receive a high score when its combination of scale responses differs from normal variation.

To test this, we evaluated all eight HPPF Industry cases (Table B). Our method achieves a mean proxy-mask P-AUROC of 67.1%. M4 provides a concrete structural example: its reference flange remains unchanged and the body’s cross-sectional dimensions are approximately preserved, while its axis tilts by 3.37° and its far-end center shifts by 3.67% of the bounding-box diagonal. In this case, our method reaches 81.6%.

These results suggest that our method can detect some large structural defects, such as the axis misalignment in M4. However, performance remains uneven across the eight cases, including two below 50%.

Evaluation setup. We use our standard normal-only training protocol, whereas HPPF is trained with pseudo-anomalies and uses a patch configuration adapted to this set. Official point-wise masks were not available to us, so we use distance-based proxy masks from the provided clean CAD meshes (d/D>10−5d/D > 10^{-5}, where DD is the clean CAD bounding-box diagonal). The set has no normal test scans, so we report only point-level localization. Our results are therefore indicative for these defect types and not directly comparable with the published scores.

Table B. HPPF Industry cases (P-AUROC, %). Baseline values are as reported in the HPPF paper; our row uses proxy masks.

MethodBearing CapDynamixel PlateFranka ConnectorHullM4 Tensioning SpoolTeeth Wrist ServoTeeth CarpalsWrist Pulley BaseAvg
R3D-AD———19.120.8———20.0
PO3AD21.625.318.236.55.97.85.130.918.9
ISMP53.535.851.168.460.357.949.937.251.8
HPPF67.864.375.678.182.385.989.561.275.6
Ours74.332.199.961.381.657.732.797.467.1

Dear Reviewer,

Thank you again for your time and effort in reviewing our work.

We would like to gently follow up on our rebuttal response. If you have any remaining questions or concerns, we would be glad to provide further clarification or additional details.

Thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.