acceptodds
Under review as a conference paper at ICLR 2027

FROM VARIATION TO VIOLATION: LEARNING 3D ANOMALY SHAPES FROM NORMAL SURFACES

Abstract

Existing 3D point cloud anomaly synthesis methods largely rely on hand-crafted deformation operators or predefined anomaly morphologies, limiting the diversity of synthesized anomalies. We observe that local changes such as bulges and depressions can also occur naturally across normal instances, and may become anomalous when they become stronger than what is normally observed. Based on this observation, we propose NOVA, a variation-driven framework for 3D anomaly synthesis and detection. NOVA uses local surface differences among aligned normal instances as deformation patterns and strengthens them beyond normal variation to create pseudo-anomalies. For detection, a normal-geometry branch first learns how much and in which directions normal surfaces vary, and is then frozen. Using its geometric features and reference-matching statistics, an anomaly-scoring branch learns from paired normal and deformed samples to assign higher scores to the same points after deformation. The resulting scores support anomaly detection and localization. Experiments confirm more diverse synthesized anomaly patterns. The normal-geometry branch alone surpasses the best published O-AUROC and P-AUROC results on both datasets using only normal training data. Supervision from 32 pseudo-anomalies per category further improves performance, raising P-AUROC from 92.68% to 95.99% on Real3D-AD and from 96.94% to 97.16% on Anomaly-ShapeNet. The resulting framework achieves O-AUROC scores of 91.28% and 96.37%, respectively. Code: https://anonymous.4open.science/r/NOVA-7C92/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D anomaly detection, point cloud anomaly detection, anomaly synthesis, pseudo-defect synthesis, industrial inspection, anomaly localization

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_roUaOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

NOVA generates synthetic 3D anomalies by extracting local deformation patterns from differences between aligned normal surfaces and amplifying them beyond an empirical normal-variation reference. The detector first learns local geometry and conditional covariance from normal data. These components are then frozen, and a scoring head is trained to rank deformed points above their normal counterparts. The paper evaluates the method on Real3D-AD and Anomaly-ShapeNet, with comparisons of synthesis methods, normal-only detection, training budgets, and registration sensitivity.

Critical strengths
  1. The main contribution, in my view, is a strong baseline with a well-specified geometric model. NOVA explicitly uses aligned normal references to model local variation, then learns a scorer from synthetic pairs. The normal-only model is already competitive. Adding the scoring branch improves all four metrics in Table 3 on both datasets; P-AP rises from 42.58% to 52.18% on Real3D-AD and from 73.73% to 76.82% on Anomaly-ShapeNet. This is a useful result, particularly given the small model and limited synthetic training set.

  2. The paper and supplementary code give enough detail to follow the pipeline, including the synthesis parameters, network dimensions, training schedules, calibration, and evaluation protocol. Per-category and per-seed results are also supplied. I appreciate the additional training-budget controls: they show that the effect of using more synthetic pairs cannot be separated from the number of optimizer updates. The documentation makes this a promising baseline for further work, although I have not run an independent reproduction.

Critical weaknesses
  1. I find the novelty limited. Learning a discriminator from synthetic anomalies is well established; the more distinctive parts here are the deformation patterns extracted from normal surfaces and the frozen normal-geometry model. The synthesis comparison is informative, but it also shows that the proposed construction is not consistently better. In Table 2, CASL-based synthesis outperforms NOVA synthesis on all four Anomaly-ShapeNet metrics with the detector fixed.

    It is also hard to tell which design choices account for the gain. The normal-only/full comparison does not isolate paired ranking from ordinary binary classification, or freezing from joint training. The quantitative parameter studies cover sample count, training duration, and head widths of 64 and 128, but leave several central geometric settings untested: deformation strength, region size, the normal-variation amplitude reference, and the normal-model geometry scale. Figure 4 only shows what different strengths look like. These omissions matter because they determine how normal variation is turned into synthetic anomalies.

  2. My main concern is the dependence on alignment and on how well the four normal references cover test-time shape variation. Appendix D.2 reports both average changes and large individual losses: the maximum observed category P-AP drops are 13.51 percentage points for Chicken and 29.99 points for cap4, each for one residual and head-training seed. What is missing is the relationship between error size and performance. At what point does detection start to deteriorate? How common are these large drops? What happens in the cases with residual rotations close to 180 degrees?

    This experiment replays transforms estimated from training normals while keeping the synthetic data and trained models fixed. It does not test errors introduced during synthesis, or separately measure the effect of genuine normal-shape differences beyond rigid pose. Cross-instance registration partly captures such differences, but mixes them with alignment effects. The current evidence leaves open how well NOVA handles valid normal shapes outside the variation represented by its reference set.

Despite these concerns, I lean toward acceptance. The performance and the detailed, usable implementation give the paper practical value. I would need stronger evidence for the individual design choices and the limits of its robustness to recommend a clear accept.

Questions for authors
  1. Could you plot the change in detection performance against residual rotation and normalized displacement? Please show the individual category/residual results as well as the trend, and report the frequency of P-AP drops above, for example, 5 and 10 percentage points. Please specify what is counted in those frequencies. I would also like to see the detection results and anomaly maps for the near-180-degree cases. Since some objects may be symmetric, a surface-alignment measure alongside the rotation angle would help interpret these cases.

  2. How does NOVA behave when a valid normal test shape differs from the four references beyond rigid pose? An analysis of false positives and detection performance as reference-shape coverage changes, with alignment accounted for, would address this concern. Have you also measured whether misalignment between training normals affects the synthetic anomalies and the detector trained on them?

  3. How does binary classification compare with paired ranking when the normal features, synthetic pairs, head, and training budget are held fixed? Separately, what is the effect of freezing versus updating the normal branch? These comparisons would help explain the gain beyond showing that the complete scoring branch works.

  4. How sensitive are the results to the added deformation amplitudes, the 512-point region size, and the quantile used for the normal-variation amplitude reference? What supports the chosen normal-model geometry scale and refinement settings? A targeted sensitivity study of these settings would help establish how much tuning the method needs.

For future work, I would be particularly interested in how far this baseline can be simplified. The head already has only about 19,000 parameters, and eight pairs per category retain much of the performance in Table 8. But the width study only compares 64 with 128. Could a linear or narrower head, a simpler deformation rule, and fewer than eight pairs achieve similar results under matched training budgets? Showing the trade-off between accuracy, model size, synthesis cost, and sample count would make the practical contribution more compelling. This is a suggestion for extending the work, not a condition for my recommendation.

4 comments
AuthorsOpenReview ↗

Response to Reviewer roUa (1/4)

Thank you for the detailed and thoughtful review, and for stating so clearly what evidence would support a clearer recommendation. We ran every analysis in your four questions, and also the simplification study you suggested. The main outcomes are:

  • Design choices. Freezing the normal-geometry branch is worth 11.2 P-AP points on Real3D-AD, and the gain from paired synthetic supervision holds with BCE as well as with the ranking loss (Q3). Patterns taken from normal surfaces add 4.5–7.4 points over a Gaussian pattern on Real3D-AD.
  • Tuning. The strength settings change P-AP by at most 0.73 points. The geometry scale and the region size are the two settings to check (Q4).
  • Robustness. Large drops under alignment error are rare (2–3% of cases lose more than 5 points), the near-180° cases change P-AP by at most 3.3 points on average, and synthesis is stable under misaligned references (Q1, Q2).
  • Simplification. Two pairs per category and a head of about 4,000 parameters keep the accuracy (last comment).

Tables R1 and R2 are in the general response; all figures are in Appendix F of the revised PDF.

Novelty, and CASL on Anomaly-ShapeNet

We agree that training a discriminator on synthetic anomalies is well established. Our claims concern the two parts you identify, and the new controls show where each one holds.

Patterns from normal surfaces. On Real3D-AD they exceed a Gaussian pattern by 4.5–7.4 P-AP points under all three strength rules, consistently across seeds and in 11–12 of 12 categories (Table R1). They also exceed a Perlin pattern placed in the same regions by 0.4–1.3 points (see our reply to yby6 for the comparison with Perlin-noise synthesis). On ASN, where normal shapes differ much less, they give no advantage: CASL and the procedural patterns are 0.5–1.3 points higher on average. For the procedural patterns, most of this difference comes from three categories (bucket0, bottle1, bottle3; Fig. F7). This agrees with the statement in the submission that no strategy is best on both datasets, and we will state the scope explicitly: the pattern helps on real scans.

Frozen normal geometry with paired supervision. Freezing is worth 11.2 P-AP points on Real3D-AD, and BCE matches the ranking loss (Q3 below). These controls support the second contribution: paired synthetic supervision applied to preserved normal-geometry features. The result does not hinge on the specific loss.

Q3. BCE vs. paired ranking; frozen vs. joint training

Table D. Loss and freezing controls (%). Normal features, synthetic pairs, head, displacement weights and budget are identical within each comparison.

DatasetSettingO-AUROCP-AUROCP-AP
Real3D-ADRanking, frozen (submitted)91.2895.9952.18 ± 0.74
Real3D-ADBCE, frozen91.4196.0852.89 ± 0.51
Real3D-ADRanking, joint86.6093.7641.00 ± 0.79
ASNRanking, frozen (submitted)96.3797.1676.82 ± 0.17
ASNBCE, frozen96.3797.0976.79 ± 0.08
ASNRanking, joint96.2697.4676.26 ± 0.09
  • Loss. Point-wise BCE on the same pairs performs comparably: 0.71 points higher on Real3D-AD and equal on ASN. We will report BCE alongside the ranking loss.
  • Freezing. Both runs start from the same normal model and head initialization; the joint run also updates the normal branch with the anomaly objective. Freezing is higher in every seed, by 11.18 points on Real3D-AD and 0.56 on ASN. On Real3D-AD, joint updates reduce the normal branch's own P-AP from 42.58 to 33.89. On ASN, joint training gives 0.30 points higher P-AUROC at lower P-AP.
AuthorsOpenReview ↗

Response to Reviewer roUa (2/4)

Q4. Sensitivity, and the basis for the settings

Table E. One setting changed at a time (P-AP, %).

SettingReal3D-ADASN
Submitted setting52.18 ± 0.7476.82 ± 0.17
Increments Δ × 0.5 / × 252.21 ± 1.49 / 52.91 ± 0.4676.94 ± 0.06 / 76.86 ± 0.24
Amplitude-reference quantile 95% / 99.9%52.45 ± 1.03 / 51.95 ± 0.8376.83 ± 0.18 / 76.88 ± 0.21
Region size 256 / 1,024 points53.52 ± 0.21 / 50.60 ± 0.8676.82 ± 0.38 / 76.84 ± 0.16
Geometry scale σ₀ × 0.5 / × 2 (retrained; matched baseline 52.01 ± 0.45)44.09 ± 0.59 / 51.15 ± 1.0176.53 ± 0.41 / 76.37 ± 0.11
Normal-only refinement removed (Real3D-AD) / added (ASN)48.48 ± 1.3677.02 ± 0.14
  • Low sensitivity: the increments and the quantile (at most 0.73 points on either dataset).
  • Moderate sensitivity on Real3D-AD: region size. Regions of 256 and 128 points are 1.34 and 1.42 points better than our 512-point default (53.52 and 53.60), while 1,024 points is 1.59 worse and 64 points becomes unstable across seeds (Appendix F, Table K). The submitted 512-point setting is therefore not the most favourable one for NOVA. We keep it as submitted and report the others alongside it.
  • High sensitivity on Real3D-AD: the geometry scale when halved (−7.9 points); doubling costs 0.9. σ₀ is set from normal data (the 75th percentile of nearest-reference distances on Real3D-AD) and is reused unchanged on ASN.
  • Refinement adds 3.5 points on Real3D-AD with the head (7.7 for the normal branch alone). On ASN it gives +0.20 P-AP and −0.15 O-AUROC; the ASN model in the paper is trained without it.

ASN moves by less than 0.5 points under every change. The practical guidance is that σ₀ and the region size are the two settings to check on a new dataset.

AuthorsOpenReview ↗

Response to Reviewer roUa (3/4)

Q1. Alignment-error size vs. performance; frequency of large drops; near-180° cases

Figs. F1–F2 plot the P-AP change for every category/residual combination (108 on Real3D-AD, 360 on ASN, each averaged over three head-training seeds) against rotation angle, displacement of corresponding points, and two-way nearest-surface distance, with binned medians. Fig. F3 gives the full category × residual matrix.

When does detection deteriorate? Surface distance is the informative axis, and the pattern is a tail of isolated failures on top of a flat median.

  • Below a surface distance of 0.1%D, no case loses more than 0.49 points on Real3D-AD (93 cases) or 1.35 points on ASN (167 cases).
  • Between 0.1% and 1%D the median change stays at 0.0 to −0.1 points, and a tail appears: 1 of 14 Real3D-AD cases and 11 of 193 ASN cases lose more than 5 points.
  • Above 1%D we have a single case (Chicken), which loses 12.7 points.

Table F. Frequency of large P-AP drops.

DatasetUnit countedDrop > 5 pointsDrop > 10 points
Real3D-ADcategory × residual, mean of three seeds2 / 108 (1.9%)1 / 108 (0.9%)
Real3D-ADcategory × residual × seed6 / 324 (1.9%)3 / 324 (0.9%)
ASNcategory × residual, mean of three seeds11 / 360 (3.1%)5 / 360 (1.4%)
ASNcategory × residual × seed33 / 1,080 (3.1%)17 / 1,080 (1.6%)

Each category contributes nine residual transforms (three source scans × three residual types), estimated on training normals and replayed on all test objects of that category.

Near-180° cases. There are 24 cases with rotation ≥ 170° (5 on Real3D-AD, 19 on ASN). Their surface distance is small, 0.31–0.40%D and 0.25–0.47%D, and the mean P-AP change is −1.40 to +0.60 and −3.29 to +0.91 points. The rotated object lies close to the reference surface, which the angle alone does not show. The largest single-seed losses occur at small angles: 13.51 points for Chicken at 5.2° and 29.99 for cap4 at 3.3°. Fig. F4 lists all 24 cases with angle, surface distance and four metrics; Fig. F5 shows anomaly maps before and after the transform on a shared colour scale for the worst near-180° case of each dataset.

Q2. Normal shapes beyond the reference set; misalignment between training normals

Shape coverage with alignment accounted for. For each normal test object (604 on Real3D-AD, 600 on ASN) we compute the remaining two-way surface distance to its closest reference after local trimmed ICP, group objects into quartiles within each category, and measure the false-positive rate of the unchanged detector at its calibrated threshold (Fig. F6).

Normal-object FPR (%)Q1 (closest)Q2Q3Q4 (farthest)
ASN3.333.123.3310.28
Real3D-AD27.7820.3720.3024.73

On ASN the false-positive rate triples for the quartile of normal shapes farthest from the references. On Real3D-AD the rate is 20–28% in every quartile and shows no monotonic relation to this distance. The threshold is calibrated on the four training normals and their scan-perturbed copies; the reported AUROC and AP do not depend on it.

Fewer references. Removing any one of the four references lowers P-AP by 1.2–6.2 points on Real3D-AD and 0.1–0.8 on ASN, and O-AUROC by 0.1–3.5 and 0.2–0.4. At the same time, the normal-object FPR of the three-reference sets is lower: 7.7–9.2% against 23.3% on Real3D-AD, and 1.9–2.9% against 4.7% on ASN. Each reference set has its own recalibrated threshold, so these rates are measured at different operating points; the threshold-free metrics above show that detection is weaker with fewer references.

Misalignment between training normals. We measured this directly (Table R2 in the general response): references are perturbed before synthesis, and patterns, pairs and the head are all regenerated. Up to 1° rotation and 0.5%D translation, P-AP stays within 0.41 points on Real3D-AD and within 0.10 on ASN; a 3° rotation costs 2.10 points on Real3D-AD and leaves ASN unchanged (76.96 against 76.90); all ASN settings stay within 0.10 points.

AuthorsOpenReview ↗

Response to Reviewer roUa (4/4)

Suggested extension: how far can the baseline be simplified?

We ran this study because it bears directly on the practical value you describe. All configurations receive the same 256 updates per category (2 / 4 / 8 / 32 pairs train for 128 / 64 / 32 / 8 epochs). Appendix F reports all 48 configurations (Figs. F9–F11).

Table G. Head size, pair count and cost (Real3D-AD / ASN).

Synthesis + headPairs per categoryHead parametersP-AP (%)Head inference per shape (ms)Generation per category (ms)
NOVA + MLP-64 (submitted)3219,201 / 19,39352.18 / 76.820.31 / 0.29110.1 / 295.1
NOVA + MLP-64219,201 / 19,39352.71 / 77.010.31 / 0.297.0 / 16.1
NOVA + MLP-1624,033 / 4,08152.44 / 76.530.23 / 0.217.0 / 16.1
NOVA + linear2234 / 23749.05 / 75.450.05 / 0.067.0 / 16.1
Gaussian + MLP-64219,201 / 19,39345.01 / 77.350.31 / 0.292.2 / 5.8

Timings are on one A100 from precomputed features; generation time starts after reference differences are prepared (CPU for Real3D-AD, GPU for ASN).

  • Fewer pairs. Two pairs per category match 32 pairs under the same update budget.
  • Narrower head. A width-16 MLP with a fifth of the parameters stays within 0.3 points. A linear head loses 3.1 points on Real3D-AD and 1.4 on ASN.
  • Simpler deformation rule. Using the mean difference to the references directly gives 16.55 / 70.48 P-AP, and a single-reference difference gives 25.26 / 72.38. Rescaling the mean difference to NOVA's RMS displacement recovers 51.99 / 76.54. This rescaled rule borrows its amplitude from NOVA; with the amplitude controlled, a simple average of normal-surface differences is a workable pattern.

Together, these results show that the baseline remains effective with a head of about 4,000 parameters and two synthetic pairs per category.

Summary against the two points you raised

Design choiceEvidence
Normal-derived pattern+4.5 to +7.4 P-AP over Gaussian on Real3D-AD; no advantage on ASN
Amplitude reference from normal data+1.3 on Real3D-AD; a dataset-level reference suffices
Frozen normal branch+11.2 on Real3D-AD, +0.6 on ASN
Loss on paired samplesRanking and BCE perform comparably
Region size, σ₀The two sensitive settings on Real3D-AD
Robustness limitMeasured boundary
Test-time misalignmentNo loss above 1.4 points below 0.1%D surface distance; isolated large losses above it (2–3% of cases over 5 points)
Misalignment during synthesisStable to 1° + 0.5%D on Real3D-AD and at every tested level on ASN; −2.1 points at 3° on Real3D-AD
Reference coverageASN false positives rise from 3% to 10% for the farthest quartile of normal shapes; one fewer reference costs up to 6.2 P-AP points

Thank you again for the constructive questions and for the simplification idea, which turned out to be one of the most useful additions to the paper. We would be glad to provide further analyses during the discussion period.

Reviewer_yby6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for anomaly synthesis based on learning the variations in surface normals of normal samples. It then uses pairs of normal-synthesized samples to train an anomaly scoring module. Experiments demonstrate very good performance on two datasets.

Critical strengths

The method is motivated based on a simple insight: normal samples exhibit variations in their surface normals which can be learned to detect anomalous samples which exhibit larger variations. This is an interesting insight which applies well to applications which involve natural samples (e.g. food, fruit, etc.) While the method is relatively simple it is quite effective and the results show substantial performance gains over previous synthesis methods.

Critical weaknesses

Although the insight that motivates the method is interesting it may not be generalized. In certain application settings (e.g. manufacturing) normal samples can be identical and may not exhibit such surface variations. This can be discussed as a limitation. The description of deformation strength and how it is controlled is not clear. This is important because it determines how much deformation is considered anomalous. if the deformation strength is hand engineered (which seems to be the case) then it most likely wont generalize to other datasets. In the figures that show point clouds, the point size is too small which makes it hard to discern normal/deformed surfaces.

Questions for authors

is the deformation strength learned or hand engineered? how it can be adapted to new datasets/applications? How does anomaly synthesis compare to OOD synthesis methods like pseudo unknown embedding (https://arxiv.org/abs/2411.18207) and perlin noise (https://arxiv.org/abs/2604.09232)?

5 comments
AuthorsOpenReview ↗

Response to Reviewer yby6 (1/3)

Thank you for the encouraging assessment and for the practical questions about applicability. Briefly, our answers are:

  • Strength. The deformation strength is computed from normal data and needs no anomalous samples. Changing its settings over a wide range moves P-AP by at most 0.73 points (W2 / Q1).
  • Other synthesis methods. We ran Perlin Raise under our detector. It transfers well to object surfaces, and our best object-level result on Real3D-AD (92.60 O-AUROC, with 54.34 P-AP) comes from training on NOVA and Perlin Raise samples together (Q2).
  • Limitation. We agree with your point about near-identical normal samples and will discuss it in the main text (W1).
  • Figures. We re-rendered the point clouds with larger points (W3).

Table R1 is in the general response; full results are in Appendix F of the revised PDF.

W1. Normal samples that are nearly identical (e.g., manufacturing)

We agree, and we will move this discussion into the main text; at present it appears only in Appendix E. NOVA takes its patterns from measurable differences between normal instances, so its benefit depends on how much the normal samples differ. The new experiments locate this boundary.

  • Scanned physical objects differ enough. Real3D-AD was built for high-precision industrial inspection and consists of high-resolution scans of physical objects. Its normal-variation reference has a median of 0.61 local point spacings across categories, and patterns extracted from these differences outperform a Gaussian pattern by 4.5–7.4 P-AP points under every strength rule (Table R1).
  • With less normal variation, the benefit shrinks. On ASN the reference is less than half as large (median 0.27 point spacings). There the normal-derived pattern performs on par with the Gaussian pattern, within about 0.6 points on average, and the detector remains effective (96.37 O-AUROC, 76.82 P-AP).
  • With exact duplicates, the difference field is zero and the same pipeline accepts a procedural pattern in its place. The Gaussian rows of Table R1 are this substitution.

One point of terminology that our text should have made clearer: the variation NOVA uses is the positional offset between aligned normal surfaces. Surface normals give the direction along which this offset is measured and along which points are displaced. We will make this explicit in Sec. 3.2.

W2 / Q1. Is the deformation strength learned or hand-engineered, and how is it adapted?

It is computed from normal data plus three fixed increments. Each point moves by (b + Δ) · ℓ along its surface normal, scaled by the pattern, where

  • ℓ is the local point spacing measured on the object, which sets the unit;
  • b is a statistic of the normal training data: the 99th percentile over regions of the 95th-percentile offset between normal surfaces;
  • Δ ∈ {0.5, 1.0, 1.5} are fixed increments, identical on both datasets.

Adapting to a new dataset requires recomputing ℓ and b from its normal samples. No anomalous samples are involved.

Table B. Sensitivity to the strength settings (P-AP, %).

SettingReal3D-ADASN
Submitted setting52.18 ± 0.7476.82 ± 0.17
Increments halved52.21 ± 1.4976.94 ± 0.06
Increments doubled52.91 ± 0.4676.86 ± 0.24
Outer quantile 95% (submitted: 99%)52.45 ± 1.0376.83 ± 0.18
Outer quantile 99.9%51.95 ± 0.8376.88 ± 0.21
One reference for the whole dataset52.17 ± 0.9776.89 ± 0.15
Increments only, no normal-data reference50.88 ± 1.2476.81 ± 0.19

Changing the increments over a fourfold range, or the quantile from 95% to 99.9%, moves P-AP by at most 0.73 points. Removing the data-derived reference costs 1.30 points on Real3D-AD, and a single reference for the whole dataset performs the same as per-category references. We will rewrite the paragraph "Controlling deformation strength" around the formula above and add Table B.

W3. Point size in the figures

We re-rendered examples with larger points, a shared camera and colour scale for each normal/deformed pair, and a locator showing the region on the full object (Fig. F8 in the revised PDF). Figs. 1, 4 and 6 will use the same rendering, with the same objects and deformations as in the submission.

AuthorsOpenReview ↗

Response to Reviewer yby6 (2/3)

Q2. Comparison with pseudo unknown embedding and Perlin noise

Thank you for pointing us to these two works. We will cite and discuss both.

Pseudo unknown embedding (arXiv:2411.18207). This work extends an open-vocabulary 2D detector to the open world. Its pseudo unknown embedding is a single class embedding, constructed in the text-embedding space from a generic object prompt and the embeddings of the known classes, and used at inference to label unknown objects. It generates no samples and no geometry, so it cannot provide deformed point clouds or point-level labels to train our detector, and a direct experimental comparison is not possible. A 3D version would need a point-cloud encoder aligned with text and a rule for point-level localization; these design choices would be ours, so the result would not represent the published method. At a conceptual level the idea is related to ours: both infer the unknown from statistics of the known (known-class embeddings there, normal-surface variation here). We will discuss it in Related Work as synthesis in embedding space, complementary to synthesis in coordinate space.

Perlin noise (arXiv:2604.09232). Perlin Raise produces geometry, so we could run it under our fixed detector with 32 pairs per category and three head-training seeds. We used the settings of the paper (target ratio 0.3, strength 0.4, radius 0.75–1.5, positive lift, largest connected cluster) and adapted it from road scenes to object surfaces by expressing its lengths in units of the local neighbourhood size. We also ran a second comparison in which a Perlin pattern is placed in the same regions as NOVA, with the same boundary taper and strength rule, so that only the pattern differs. There are three findings.

  1. Perlin Raise transfers well to object surfaces, and it combines well with NOVA. Perlin Raise alone gives 92.34 O-AUROC and 54.30 P-AP on Real3D-AD, which is 2.12 P-AP points above the submitted NOVA setting (91.28 and 52.18); on ASN it is 0.54 points above. Training on 16 NOVA pairs and 16 Perlin Raise pairs per category gives 92.60 O-AUROC, the highest value our detector has reached on Real3D-AD, and 54.34 P-AP, which matches Perlin Raise alone within seed variation.
  2. With regions and strength held fixed, the normal-derived pattern is higher than the Perlin pattern on Real3D-AD. The difference is 1.27, 0.99 and 0.44 P-AP points at region sizes of 512, 256 and 128 points (Table K in the next comment).
  3. The size of the deformation matters. Perlin Raise displaces a median of 33 points per sample, almost all outward; NOVA displaces 272 points in both directions, with a similar peak displacement (1.8 and 1.9 local point spacings). This is what led us to vary the region size of NOVA (next comment).

Table C. Perlin comparison under the fixed NOVA detector (%).

DatasetSynthesisO-AUROCO-APP-AUROCP-AP
Real3D-ADNOVA91.2889.5995.9952.18 ± 0.74
Real3D-ADPerlin pattern, NOVA regions and strength91.2789.8895.6750.91 ± 0.44
Real3D-ADPerlin Raise, published settings92.3490.0796.6054.30 ± 0.34
Real3D-ADNOVA + Perlin Raise, 16 pairs each92.6090.3396.5654.34 ± 0.32
ASNNOVA96.3798.0097.1676.82 ± 0.17
ASNPerlin pattern, NOVA regions and strength96.6198.0997.6177.86 ± 0.18
ASNPerlin Raise, published settings96.2097.9497.4077.36 ± 0.10

On ASN the generators lie within about one point of each other (76.82–77.86 P-AP), in line with the Gaussian control. As there, the differences come mainly from three categories (bucket0, bottle1, bottle3); Fig. F7 shows all 52 categories.

AuthorsOpenReview ↗

Response to Reviewer yby6 (3/3)

Q2 (continued). Region size, and what we will change

Because Perlin Raise uses much smaller deformations, we varied the region size and the increments of NOVA on Real3D-AD and repeated the matched Perlin-pattern comparison at 512, 256 and 128 points.

Table K. Region size and increments on Real3D-AD (P-AP, %). The Perlin column uses the same regions and strength rule as the NOVA column to its left. Appendix F gives the full table with all four metrics, including a 64-point setting at which NOVA becomes unstable across seeds.

Region (points)NOVA, increments ΔPerlin pattern, increments ΔNOVA, increments 2Δ
512 (submitted)52.18 ± 0.7450.91 ± 0.4452.91 ± 0.46
25653.52 ± 0.2152.53 ± 0.6953.54 ± 0.27
12853.60 ± 0.2053.17 ± 0.3653.18 ± 0.38

Two things follow. First, the submitted 512-point setting is not the most favourable one for NOVA: with regions of 128–256 points it reaches 53.2–53.6 in all four runs, 1.0–1.4 points above the submitted result and with a smaller spread across seeds. This recovers about two thirds of the difference to Perlin Raise (54.30). Second, the normal-derived pattern stays ahead of the Perlin pattern at each of these three region sizes.

In the revised paper we will add Perlin Raise and the matched Perlin pattern to the comparison of Table 2, report Table K, cite both papers, and discuss deformation size as a design factor. We are grateful for the suggestion: the comparison gave us a stronger overall result and a clearer picture of what the normal-derived pattern contributes.

Reviewer_6qeiOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NOVA, a variation-driven framework for 3D anomaly synthesis and detection. The method extracts deformation patterns from surface differences among aligned normal objects and controls their strength relative to observed normal variation, generating diverse pseudo-anomalies without predefined deformation operators. For detection, a normal-geometry branch is first trained on normal samples and then frozen, while an anomaly-scoring branch learns from paired normal and deformed samples. Experiments on Real3D-AD and Anomaly-ShapeNet demonstrate strong object-level detection and point-level localization performance.

Critical strengths

S1. Generating pseudo-anomalies from geometric variations in normal objects offers an interesting alternative to manually designed deformation operators. Using normal variation to guide both deformation patterns and strength is well motivated.

S2. The proposed framework separates normal-geometry modeling from synthetic anomaly supervision. Table 3 shows that adding synthetic supervision improves P-AUROC from 92.68% to 95.99% and P-AP from 42.58% to 52.18% on Real3D-AD.

S3. The paper presents comprehensive experimental comparisons and controlled evaluations. Table 2 compares different synthesis strategies under a fixed detector, while Figure 5 and Appendix D.1 investigate the effects of synthesis quantity and training budget.

Critical weaknesses

W1. Although Table 2 compares different anomaly synthesis strategies, it does not fully disentangle the contributions of deformation patterns and strength calibration. The Gaussian control partially isolates the effect of deformation patterns by using the same regions and strength schedule as NOVA. However, the benefit of estimating deformation strength from normal surface variation remains unclear. A more systematic ablation that independently varies the deformation pattern source and strength calibration strategy would better justify the proposed design.

W2. The proposed normal-geometry branch incorporates several components, including a learned covariance predictor, Gaussian-mixture likelihood, and Mahalanobis distance. However, the paper does not provide sufficient ablation studies to isolate their individual contributions. Since the normal-only detector already achieves strong performance, further analysis of these components would help clarify which design choices are responsible for its effectiveness.

W3. The proposed anomaly synthesis strategy relies on accurate alignment between normal point clouds to extract meaningful surface differences. Although Appendix D.2 investigates test-time registration sensitivity, it does not evaluate the impact of registration errors during pseudo-anomaly generation. The authors also acknowledge this limitation. Further experiments with imperfectly aligned normal references would help establish the robustness and practical applicability of the proposed synthesis framework.

Questions for authors

See weaknesses

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6qei (1/3)

Thank you for the careful review. Each of your three points asked for a specific experiment, and we have run all three. The outcomes are:

  • W1. On Real3D-AD both elements contribute: the pattern source adds 4.5–7.4 P-AP points, and the strength estimated from normal variation adds 1.3.
  • W2. The normal-geometry branch depends on its learned local variance, its geometry scale and the normal-only refinement. The exact parametric form matters little.
  • W3. Synthesis is stable under registration errors about ten times larger than the residuals we measure on Real3D-AD, and at every level we tested on ASN.

Tables R1 and R2 are in the general response, and Appendix F of the revised PDF has the four-metric versions.

W1. Separating the deformation pattern from strength calibration

We ran the factorial you describe (Table R1): two pattern sources (normal-surface differences and an anisotropic Gaussian) crossed with three strength rules (a per-category reference b plus the increments Δ; one dataset-level reference plus Δ; Δ alone). Regions, boundary taper, increments, detector and training budget are the same in all six cells.

Pattern source. On Real3D-AD the normal-derived patterns are 5.73, 4.50 and 7.43 P-AP points above the Gaussian patterns under the three strength rules. The sign is the same in all three head-training seeds and in 11–12 of the 12 categories.

Strength estimated from normal variation. Adding the normal-derived reference to the fixed increments raises P-AP on Real3D-AD by 1.30 points with NOVA patterns (52.18 against 50.88; +2.23, +0.63 and +1.05 in the three seeds; 10 of 12 categories) and by 3.00–4.22 points with Gaussian patterns.

Per-category or dataset-level reference. The two give the same result (52.18 and 52.17), so the reference can be estimated once per dataset, which makes the method simpler. We will say so in the paper.

Anomaly-ShapeNet. Normal shapes differ much less here. The reference has a median of 0.27 local point spacings (range 0.13–0.48), against 0.61 on Real3D-AD (0.19–2.73), while the increments are 0.5–1.5. Accordingly the six cells lie within about 0.6 points of each other. Gaussian is 0.3–0.6 points higher on average because of three categories (bucket0, bottle1, bottle3), and NOVA is higher in 21–24 of the 40 categories (Fig. F7).

So on real scans both elements matter, and the pattern source contributes several times more than the amplitude reference. We will add Table R1 to Sec. 4.3 in place of the sentence that these effects "remain to be established".

AuthorsOpenReview ↗

Response to Reviewer 6qei (2/3)

W2. Which components of the normal-geometry branch matter

We changed the branch one component at a time (Table A). Three things account for its effectiveness.

  1. The learned local variance. When the covariance predicted by the branch is replaced with a fixed value in the score and in the head inputs, P-AP with the scoring head falls from 52.18 to 31.85 on Real3D-AD and from 76.82 to 73.47 on ASN. The scoring head relies on this learned uncertainty.
  2. Normal-only refinement, on Real3D-AD. From the same starting model, refinement raises normal-only O-AUROC / P-AP from 79.25 / 34.90 to 88.54 / 42.58, and P-AP with the scoring head from 48.48 to 52.01.
  3. The geometry scale σ₀. Halving it lowers P-AP with the scoring head on Real3D-AD by 7.9 points; doubling it lowers P-AP by 0.9.

The exact parametric form is less critical. A single learned variance for all directions, a likelihood fitted to the nearest reference only, or a Euclidean distance in place of the Mahalanobis statistic all stay within 0.7 points of the submitted model on Real3D-AD. On ASN the submitted model keeps about 0.5 points higher normal-only O-AUROC than the single-variance and Euclidean variants, while those variants have 1.7–1.9 points higher normal-only P-AP.

Table A. Normal-geometry branch variants. One normal model per variant; ± is over three head-training seeds.

DatasetVariantNormal-only O-AUROCNormal-only P-APWith scoring head, P-AP
Real3D-ADSubmitted model88.5442.5852.18 ± 0.74
Real3D-ADOne learned variance for all directions88.4942.6452.54 ± 0.18
Real3D-ADLikelihood fitted to the nearest reference only88.5542.6151.52 ± 0.44
Real3D-ADEuclidean distance replaces Mahalanobis statistic88.5442.6552.36 ± 0.72
Real3D-ADFixed covariance in score and head inputs88.5442.6531.85 ± 13.60
Real3D-ADWithout normal-only refinement79.2534.9048.48 ± 1.36
ASNSubmitted model96.1273.7376.82 ± 0.17
ASNOne learned variance for all directions95.5675.4676.49 ± 0.20
ASNLikelihood fitted to the nearest reference only96.3375.4777.07 ± 0.17
ASNEuclidean distance replaces Mahalanobis statistic95.5875.6076.81 ± 0.13
ASNFixed covariance in score and head inputs95.5875.6073.47 ± 1.20

The two retrained variants and the no-refinement row are compared with a retrained control at 52.01 ± 0.45; the other rows use the submitted model. Table A is now in Appendix F, and we will mention the single-variance form as a simpler alternative.

AuthorsOpenReview ↗

Response to Reviewer 6qei (3/3)

W3. Registration errors during pseudo-anomaly generation

We perturbed the normal references before synthesis (rotations of 0.25°, 1° and 3°; translations of 0.1%, 0.5% and 1% of the bounding-box diagonal D; and their combinations), recomputed correspondences, patterns and amplitude references, regenerated all pairs, and retrained the scoring head. The normal model and the test set are unchanged (Table R2).

  • Real3D-AD. Every setting up to 1° and 0.5%D stays within 0.41 points of the unperturbed 52.28 (range 52.07–52.69). A 3° rotation costs 2.10 points, and 3° + 1%D costs 1.30.
  • ASN. All settings, including 3° + 1%D, stay within 0.10 points of the unperturbed 76.90.
  • Scale. A rotation of 1° or a translation of 0.5%D moves the surface by a median of 0.33–0.47%D. The registration residuals measured in Appendix D.2 have medians of 0.014–0.035%D on Real3D-AD and 0.06–0.28%D on ASN. Synthesis is therefore stable at misalignments about ten times the typical measured residual on Real3D-AD, and at every level we tested on ASN (up to 0.76%D).

One note: the 3° setting revealed a numerical instability of the pattern normalizer in sparsely matched regions, which we corrected for this experiment. The unperturbed results are essentially unchanged (52.18 → 52.28; 76.82 → 76.90); details are in the general response.

This study is now in Appendix F, next to the test-time study of Appendix D.2, and the Limitations sentence that synthesis-stage sensitivity "remains to be established" will be replaced by these measured limits.

Thank you again for these three suggestions; they have made the paper more complete. We would be glad to run further controls during the discussion period.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response: what the new experiments show

We thank all three reviewers for their careful and constructive reviews. We have run every analysis the reviews asked for. The complete results are in a new Appendix F of the revised PDF (marked in blue), and the numbers in the main text are unchanged. This note summarizes the outcome; the individual replies give the details.

Main findings

  1. Both design elements the reviewers singled out are supported on Real3D-AD. Deformation patterns taken from normal surfaces raise P-AP by 4.5–7.4 points over a Gaussian pattern (Table R1), and by 0.4–1.3 points over a Perlin pattern placed in the same regions with the same strength. Freezing the normal-geometry branch is worth 11.2 points over joint training.
  2. The results are stable. Changing the strength increments fourfold, or the amplitude quantile, moves P-AP by at most 0.73 points, and BCE performs like the ranking loss. Synthesis tolerates registration errors of 1° and 0.5% of the object diagonal, about ten times the residual we measure on Real3D-AD (Table R2).
  3. The baseline can be much lighter than submitted. Two synthetic pairs per category and a scoring head of about 4,000 parameters keep the accuracy of the submitted model.
  4. The detector also works with other synthesis methods. Trained on NOVA and Perlin Raise samples together, it reaches 92.60 O-AUROC on Real3D-AD, the highest value we have obtained, with 54.34 P-AP.

We also measured where the limits are, as the reviewers asked. On Anomaly-ShapeNet (ASN), where normal shapes differ much less, procedural patterns perform on par with ours or up to 1.3 points higher, as the submission already reports for CASL. The geometry scale σ₀ and the region size are the two settings worth checking on a new dataset.

Shared protocol. Unless a setting is varied explicitly, all new experiments cover the 12 Real3D-AD and 40 Anomaly-ShapeNet (ASN) categories, with 32 normal/deformed pairs per category, eight epochs, and three head-training seeds. The normal-geometry model and the synthesis pool are fixed across seeds. We report the mean ± SD of the category-averaged metric; P-AP is in %, and differences are in percentage points. The NOVA reference of 52.18 P-AP on Real3D-AD is the 32-pair result of Table 3 in the submission; Table 2 there uses 16 pairs per Real3D-AD category and reports 52.49.

#New experimentRequested byLocation
1Pattern source × strength rule, 2 × 3 factorial6qei W1, roUa, yby6Table R1 below
2Components of the normal-geometry branch6qei W2reply to 6qei
3Misaligned normal references during synthesis6qei W3, roUa Q2Table R2 below
4Test-time alignment error vs. performance, near-180° casesroUa Q1reply to roUa, Figs. F1–F5
5Reference coverage and false positivesroUa Q2reply to roUa, Fig. F6
6BCE vs. ranking; frozen vs. joint trainingroUa Q3reply to roUa
7Sensitivity to increments, quantile, region size, geometry scale, refinementroUa Q4, yby6replies to roUa and yby6
8Comparison with Perlin Raise and a matched Perlin patternyby6reply to yby6
9Narrower heads, fewer pairs, simpler deformation rules, costroUareply to roUa

Table R1. Pattern source × strength rule (P-AP, %). Regions, boundary taper, increments Δ ∈ {0.5, 1, 1.5}, detector and training budget are identical in all six cells.

DatasetPattern sourcePer-category reference + ΔDataset-level reference + ΔΔ only
Real3D-ADNormal-surface differences (NOVA)52.18 ± 0.7452.17 ± 0.9750.88 ± 1.24
Real3D-ADAnisotropic Gaussian46.45 ± 0.5847.67 ± 1.0343.44 ± 1.01
ASNNormal-surface differences (NOVA)76.82 ± 0.1776.89 ± 0.1576.81 ± 0.19
ASNAnisotropic Gaussian77.42 ± 0.1677.43 ± 0.2277.09 ± 0.25
AuthorsOpenReview ↗

General response (continued): synthesis under misaligned references

Reviewers 6qei and roUa asked what happens when the normal references are misaligned during synthesis. Table R2 gives the answer: on Real3D-AD, P-AP stays within 0.41 points of the unperturbed result for every setting up to 1° + 0.5%D, and on ASN all settings stay within 0.10 points.

Table R2. Normal references perturbed before synthesis (P-AP, %). Correspondences, patterns and amplitude references are recomputed from the perturbed references, all pairs are regenerated, and the scoring head is retrained. D is the bounding-box diagonal. The second column gives the median two-way nearest-neighbour RMS distance between a reference and its perturbed copy (Real3D-AD / ASN), the same surface measure used in our reply to roUa Q1.

PerturbationSurface distance (%D)Real3D-ADASN
None052.28 ± 0.7576.90 ± 0.16
Rotation 0.25°0.09 / 0.0952.43 ± 0.8476.96 ± 0.13
Translation 0.1%D0.10 / 0.1052.07 ± 1.0476.83 ± 0.08
0.25° + 0.1%D0.14 / 0.1452.33 ± 1.0176.91 ± 0.13
Rotation 1°0.33 / 0.3452.15 ± 0.9577.00 ± 0.18
Translation 0.5%D0.44 / 0.4752.69 ± 0.9376.82 ± 0.13
1° + 0.5%D0.46 / 0.4752.35 ± 0.3876.97 ± 0.17†
Rotation 3°0.60 / 0.5750.18 ± 0.1276.96 ± 0.15†
Translation 1%D0.65 / 0.6752.32 ± 1.5576.87 ± 0.14†
3° + 1%D0.84 / 0.7650.99 ± 1.6976.93 ± 0.29†

† In these four ASN settings, 1, 9, 15 and 24 of the 1,280 sampled regions had no usable pattern (no point with a valid match in all three references, or an all-zero pattern). We replaced only those regions with the next candidate region and kept everything else unchanged.

A note on the normalizer. At 3° on Real3D-AD, the pattern normalizer Z of Sec. 3.2 became unstable in a few regions with very sparse matches, so for this experiment we compute Z over the commonly valid support (Appendix F.2). With unperturbed references this changes P-AP from 52.18 to 52.28 on Real3D-AD and from 76.82 to 76.90 on ASN. Table R2 uses this version with its own unperturbed baseline; all other tables use the definition in the submission.

Additions to the paper. The new tables and figures are in Appendix F. In the final version we will also (1) report Table R1 in Sec. 4.3 and describe the amplitude reference as a dataset-level statistic, (2) report the Perlin Raise comparison alongside Table 2, update the discussion of Real3D-AD accordingly, and discuss both papers suggested by Reviewer yby6, and (3) discuss near-identical normal samples, misalignment, and reference coverage as limitations in the main text, with the measured limits.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.