acceptodds
Under review as a conference paper at ICLR 2027

ANYNORMAL: LET EVERY POINT CHOOSE ITS NORMAL

Abstract

In few-shot 3D anomaly detection, only a handful of normal examples are available to characterize normality, even though normal objects can differ in local shape and sampling. This raises two questions: how should a geometric difference be judged against the variation among normal objects, and which normal example should be used to assess each test region? We propose AnyNormal. It first compares the normal examples with one another to estimate the range of normal geometric variation. At test time, each query point is matched independently to every aligned normal reference. For each reference, Euclidean, point-to-plane, and normal-angle differences are measured relative to the variation observed among normal examples, at both the point and local-neighborhood levels, and averaged into a score. The minimum score across references is used for each point, allowing different regions of the same object to be assessed against different normal examples. No scoring network is trained: a parameter-free mean of these calibrated geometric features achieves 90.3/95.1%, 94.5/96.2%, and 89.0/87.0% object/point AUROC on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD, respectively. In a transfer experiment, a learned scorer trained on only four normal scans from a single Real3D-AD category reaches essentially the same accuracy across all three datasets without further weight updates. A compressed configuration achieves 85.9/82.4%, 90.1/88.0%, and 79.3/71.4% object/point AUROC on the three datasets, respectively, while reaching 69.74 FPS in continuous-stream evaluation on an RTX 4090. Code for the parameter-free method is available at https://anonymous.4open.science/r/AnyNormal/.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_iQCcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AnyNormal, a training-free method for few-shot 3D anomaly detection. Its two ideas are normal-variation calibration, which calibrates Euclidean, point-to-plane, and normal-angle residuals against the variation among normal objects, and per-reference scoring, which scores each reference before selecting via a pointwise minimum rather than pooling references first. A parameter-free fixed mean of the calibrated features serves as the scorer. Across Real3D-AD, Anomaly-ShapeNet and MulSen-AD it matches or exceeds learned baselines, and a scorer trained on one category transfers across datasets without further updates.

Critical strengths
  1. The core contribution is a well-supported negative result: for few-shot 3D anomaly detection, a scorer with no learned parameters is enough.
  2. Table 5 isolates calibration, local context, and per-reference scoring under matched conditions with confidence intervals, and reports cases where a cue is not best.
Critical weaknesses
  1. Novelty is at the level of insight and recombination, not a new mechanism. The components are individually standard. The contribution is the score-then-select framing plus calibration against normal variation. The isolated gains are modest.
  2. As the authors acknowledge, appearance-only defects and reference-absent normal variation are out of scope, and misalignment can be mistaken for a defect. Numerous constants are fixed with appendix justification but without sensitivity analysis, leaving robustness partly unverified.
Questions for authors
  1. On MulSen-AD the main result uses the full normal training set rather than four-shot support. What accuracy does the method reach on MulSen-AD under the same four-shot setting as the other two benchmarks? This directly affects how firmly the few-shot claim can be stated.
  2. How do the results change if the main constants, such as the percentile levels, the clipping constant, and the orientation threshold, are varied?
  3. The analysis that a learned scorer collapses to the fixed mean is derived for a monotone positive-weight MLP. Does the same collapse occur with an unconstrained scorer, or is this conclusion specific to the monotonicity constraint?
  4. Scoring uses a pointwise minimum across references, so a point is flagged only when every reference scores it high. How does accuracy change as the number of references grows, and is there evidence the minimum does not erode recall?
4 comments
AuthorsOpenReview ↗

Response to Reviewer iQCc (1/2): four-normal support and sensitivity

iQCc — 1/2

Response to Reviewer iQCc

Thank you for highlighting the training-free result. Normal references provide the shape information and the scale for judging deviations; the same fixed scoring rule can then be used across categories. Our contribution is to show which comparisons make this possible: measuring geometric differences against normal variation and scoring each reference before selection.

With neighborhood averaging fixed, calibration adds 5.48/1.35/4.10 point-AUROC points on Real3D-AD/Anomaly-ShapeNet/MulSen-AD, using four references on the first two and full support on MulSen-AD. At four scored matches, per-reference scoring adds 0.68/1.50 point-AP points on the first two datasets at similar CPU latency.

The scorer controls help explain why this works. Starting from the same ranking, unconstrained normal-only training can reverse it, while positive-weight scoring stays close to fixed averaging (Q3). The useful information is already present in the calibrated geometric features. A fixed mean uses it without having to fit scoring weights for each category.

Q1. Strict four-normal MulSen-AD

We evaluated all 15 categories and 644 objects for each of three support selections, using the same four scans to fit alignment, normalization, calibration and the reference bank with one fixed registration profile.

Table I1. MulSen-AD support comparison. Object/point AUROC (%); mean ± support-selection SD. Tables 1 and 16, pp. 5 and 20.

Normal supportObject AUROCPoint AUROC
Strict four-normal85.51 ± 1.9681.67 ± 0.64
Original full-normal88.9687.04

The strict four-normal point AUROC exceeds the published baseline values listed in Table 1(b), including Simple3D (80.3%) and MAD-3D (80.1%), under their respective published protocols. The table reports our two support regimes separately.

Q2. Parameter sensitivity and geometric conditions

We varied each parameter individually and retained the original defaults.

Table I2. Point-AUROC ranges (%). Main support settings; MulSen-AD full support. Table 17, p. 20.

Tested parameterReal3D-ADAnomaly-ShapeNetMulSen-AD
Five percentile pairs94.91–95.2096.12–96.3486.71–87.42
Clipping: 5, 10, 20, 4095.05–95.0696.24–96.2686.98–87.05
Orientation: 0.25, 0.5, 0.7595.06–95.0696.07–96.3287.03–87.04
Neighborhood: 8, 16, 3294.41–95.8695.66–96.7686.65–87.41

Percentile pairs are (90,99), (95,99), (95,99.5), (95,99.9) and (97.5,99.5). Percentile and clipping effects are small over these ranges. Orientation affects Anomaly-ShapeNet point AP more: 68.11/65.20/61.81% at thresholds 0.25/0.5/0.75.

We also tested the alignment and normal-shape coverage concerns raised in your review (Tables 21 and 23–24). On 12 objects, rerunning registration recovers geometry within 1% of the diagonal of the normal-reference bounding box in 107/108 perturbations, relative to the registration output without the added perturbation. At equal reference counts, geometrically diverse normal supports improve point AUROC by 1.46/0.93/2.91 points on Real3D-AD/Anomaly-ShapeNet/MulSen-AD.

AuthorsOpenReview ↗

Response to Reviewer iQCc (2/2): scorer constraints and recall

iQCc — 2/2

Q3. Removing the positive-weight constraint

No. Under the same normal-only objective, an unconstrained scorer can reverse the anomaly ranking. Both networks use 12→32→16→1 layers, Softplus hidden activations and 120 updates on four normal airplane scans. We test three seeds, including a model initialized with exactly the same weights as the positive-weight network.

Table I3. Point AUROC (%) before/after scorer training. Three-seed means. Table 26, p. 25.

ScorerReal3D-ADAnomaly-ShapeNetMulSen-AD
Fixed mean95.0696.2487.04
Positive-weight MLP, random initialization (untrained)95.0396.2487.01
Positive-weight MLP, random initialization (trained)95.0296.2686.98
Unconstrained MLP, same initial weights (trained)9.385.7828.79
Unconstrained MLP, random signed weights (trained)9.416.4428.02

When the two networks start with identical weights, they give the same ranking, but unconstrained training reverses it. The normal-only loss rewards low scores on normal examples without requiring larger deviations to receive higher scores. Positive weights preserve this increasing score direction. With constant initialization, hidden units remain identical: the network reduces to an increasing function of a positive-weight sum of the features. The near-equality to fixed averaging under random initialization is an experimental result. Fixed averaging retains the intended score direction without this fitting step.

Q4. Reference count and recall

ROC recall at FPR ≤5% improves as the bank grows, but keeping the old threshold causes more misses. This study fixes the original coordinates and calibration, so its four-reference baseline differs from Q1, which recomputes all stages from four scans.

Table I4. MulSen-AD reference-bank growth. Percentages, three nested orders; recall is the maximum empirical ROC recall at FPR ≤5%. FNR/FPR use the four-reference normal-test-score 95th-percentile threshold retrospectively. “All” uses all available normal training scans per category.

ReferencesPoint AUROCROC recall@5%FPRFNR / FPR (fixed threshold)
481.4459.3938.49 / 7.05
883.4060.9844.54 / 2.06
1684.9664.5147.86 / 0.75
All87.0468.3250.52 / 0.55

The larger bank improves ranking and recall at the same FPR limit. At full support, normal/anomalous score ratios to their four-reference values have medians of 0.25/0.89 across category/order pairs. Normal scores fall more, so keeping the old threshold produces fewer false alarms and more misses. The threshold should therefore be checked for the bank used at deployment.

Under the same fixed-coordinate/calibration protocol, increasing from two to four references raises point AUROC from 93.71 to 95.06% on Real3D-AD and 95.45 to 96.24% on Anomaly-ShapeNet.

AuthorsOpenReview ↗

Follow-up to Reviewer iQCc: novelty and the size of the gains (Weakness 1)

Our first response covered Q1–Q4. This note answers Weakness 1 directly.

We agree with your description: the components are standard, and the contribution is the score-then-select framing, calibration against normal variation, and the result that a parameter-free scorer is then sufficient. We see this as the strength of the paper. It shows that accurate few-shot 3D detection that generalizes does not require complex network design: one fixed rule with the same constants covers 67 categories of three benchmarks without training, and a compressed configuration runs at 20.46 ms per object with its accuracy cost reported beside it (Table 5). Three points support this.

1. The framing isolates a design choice. In memory-bank detectors the matching distance is the anomaly score, so selecting a reference and scoring it are one step. With calibrated point-to-plane and normal-angle residuals they become two steps, and their order matters: among the 64 pooled candidates of a query point, the lowest-scoring one is not the nearest one at 38%/41% of points on Real3D-AD/Anomaly-ShapeNet (pointwise features). All 13 variants in Table 25 that score before selecting outperform selecting by proximity first in point AP, on both datasets.

2. The effects are larger than the matched-budget number suggests. The 0.68/1.50 AP points we quoted is our most conservative control: its baseline already scores four candidates before selecting.

ComparisonGain
Score-then-select vs. select-then-score (Table 6b)+2.50 / +2.65 point AP on Real3D-AD / Anomaly-ShapeNet; 12/12 and 36/40 categories improve
Calibration, neighborhood averaging fixed (Table 2a)+5.48 / +1.35 / +4.10 point AUROC on the three datasets
Full method vs. Simple3D, identical inputs and references (Sec. 3.2)+19.7 / +13.1 object/point AUROC on Real3D-AD; +11.2 / +6.6 on Anomaly-ShapeNet
Four scans vs. full-support baselines, MulSen-AD (Table 1b)81.67 vs. best listed 80.3 point AUROC

3. The negative result you highlighted now has an explanation. Your Q3 led to the control we consider the most informative addition. From identical initial weights, removing the positive-weight constraint lets normal-only training reverse the ranking (point AUROC 9.4/5.8/28.8%), while positive-weight scorers stay within 0.1 points of the fixed mean (Table 26). What the scorer needs is the direction "larger calibrated deviation, higher score"; the fixed mean supplies it without training.

On Weakness 2 and Q4: Table 17 reports the sensitivity sweeps, Figure 6 (p. 22) plots accuracy, recall and threshold behavior against bank size, and the Limitations paragraph states the alignment, coverage and threshold conditions.

We would be glad to clarify anything that remains open.

AuthorsOpenReview ↗

Gentle reminder

We would be grateful to know whether the four-shot MulSen-AD results, the parameter sensitivity study and the unconstrained-scorer control address your questions, and we are happy to clarify anything further before the discussion period closes.

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

This work introduces AnyNormal, a largely parameter‑free few‑shot 3D anomaly detection method leveraging normal‑variation calibration and per‑reference scoring. It computes three normalized geometric residuals for each query‑reference pair, constructs 12‑dimensional point‑and‑neighborhood calibrated features, and produces per‑reference scores via fixed‑weight averaging. Point‑level scores take the minimum across references to enable region‑wise reference selection, while object‑level scores average the top‑1% point scores. It obtains 90.3/95.1%, 94.5/96.2%, 89.0/87.0% object/point AUROC on Real3D‑AD, Anomaly‑ShapeNet and MulSen‑AD. Additional experiments cover point AP, a 52‑category extended setting, a compressed real‑time variant, and a transferable monotone scorer trained on merely four normal scans from one Real3D‑AD category, with full ablations comparing residual cues, calibrated features, pooled matching and different scoring strategies.

Critical strengths
  1. The method has no learned scoring parameters, synthetic anomalies, or complex backbone, yet is competitive on Real3D-AD and Anomaly-ShapeNet and improves point localization substantially on MulSen-AD. This simplicity makes the contribution easy to inspect and potentially useful in practice.
  2. The paper includes confidence intervals, per-category and expanded-category results, cue and scorer studies, a speed/accuracy trade-off, and anonymous code. The near-equivalence of fixed and learned scoring supports the claim that calibrated features, rather than a flexible predictor, carry most of the signal.
Critical weaknesses
  1. Normalizing residuals by empirical normal quantiles and changing the order from pooled match-then-score to score-each-reference-then-minimize are intuitive robust matching choices. The paper does not provide a new statistical objective, guarantee, learned representation, or analysis showing when this ordering is preferable.
  2. The headline 10.56/5.55-point gain combines quantile calibration with neighborhood averaging, while per-reference scoring performs more reference-specific evaluations than pooled matching. Consequently, the paper does not establish how much improvement comes from its named innovations rather than extra context or computation. Dependence on alignment and target-normal coverage is also not stress-tested, and the main MulSen setting is not few-shot.
Questions for authors
  1. Could the authors run separate ablations for (a) quantile‑based normal‑variation calibration, (b) local‑neighborhood averaging, and (c) the combination of both? Is calibration alone responsible for most of the observed point‑AUROC improvement?
  2. Could pooled matching and per‑reference scoring be compared under matched numbers of distance evaluations, scored matches, memory footprint, and runtime? I would like to see both accuracy and end‑to‑end latency from this controlled comparison.
  3. I am curious how accuracy shifts with the count and diversity of normal references, controlled registration perturbations, missing normal modes, and alignment failure. For MulSen‑AD, could the authors also supply a four‑shot protocol aligned with the setup used for the remaining datasets.
  4. Could the authors document the full objective function, optimizer, learning rate, model‑selection rule, and random seeds used for the learned scorer? I also note that the existing transfer result should be labelled as scorer transfer with target‑category normal support, instead of detector‑level zero‑shot transfer.
  5. Since the point‑level score takes a minimum across references, adding more references may systematically depress scores for both normal and anomalous samples. How does the false‑negative rate evolve as the reference bank expands, especially when one reference locally mimics an anomalous deformation? Could the authors present performance and calibration curves plotted against reference‑bank size?
  6. Are the 95 % and 99.5 % normal‑residual quantiles estimated per‑category, per‑reference, per‑spatial‑region, or globally? How many effective samples underpin the 99.5 %‑tail estimate? Could the authors report quantile sensitivity and associated uncertainty under four‑shot sampling conditions?
7 comments
AuthorsOpenReview ↗

Response to Reviewer dXmL (1/3): calibration and matched-budget comparisons

dXmL — 1/3

Response to Reviewer dXmL

Thank you for the detailed questions. The key distinction is how we choose a normal reference. The nearest reference point need not give the lowest anomaly score once normal variation is taken into account. AnyNormal scores each reference using calibrated local geometry before making that choice, so different test regions can use different normal examples.

We tested whether the gains come from calibration, neighborhood averaging or scoring more candidates. Calibration improves localization with neighborhood averaging fixed; per-reference scoring also improves point AP at four scored candidates and similar CPU latency. Q1–Q2 report the corresponding controls.

Q1. Calibration versus local-neighborhood averaging

Calibration accounts for the larger standalone gain on Real3D-AD and MulSen-AD; local averaging contributes more on Anomaly-ShapeNet. The requested 2×2 comparison holds matching, reference selection and fixed-mean scoring constant.

Table D1. Calibration × local context. Point AUROC (%); MulSen-AD uses full support. Table 2(a), p. 7.

DatasetRawRaw + localCalibratedCalibrated + local
Real3D-AD84.5089.5793.2095.06
Anomaly-ShapeNet90.6994.9093.5996.24
MulSen-AD79.1282.9485.3887.04

On Real3D-AD/Anomaly-ShapeNet/MulSen-AD, calibration with local context fixed adds 5.48/1.35/4.10 points, with paired category-bootstrap 95% intervals [4.06,6.78]/[1.10,1.62]/[3.00,5.23]. Without context, the gains are 8.70/2.90/6.27 points. The original headline gain combines calibration and neighborhood averaging; their effects are not additive. Calibration here uses two features: how much a residual exceeds the normal level, and its magnitude relative to that level.

Q2. Candidate budgets, resource use and end-to-end latency

All methods share the reference points and leave-one-out calibration: each normal scan is compared with the merged remaining scans. Each query retrieves 64 neighbors, either from the merged references or 16 per reference. We vary how many matches receive a score.

Table D2. Candidate budget and measured CPU scoring cost. Point AP (%) covers 2,518 objects. CPU p50 values are Real3D-AD / Anomaly-ShapeNet, measured on 104 predetermined objects with three repetitions. Tables 2(b), 25 and 27, pp. 7, 24 and 26.

Scoring ruleReal3D-AD APAnomaly-ShapeNet APCPU p50 (ms)
Merged references, 1 match43.0663.37542.70 / 535.80
Merged references, 4 matches44.8964.52603.80 / 596.22
Merged references, 16 matches45.4764.62832.48 / 819.54
Merged references, 64 matches45.6164.681726.66 / 1712.23
One match per reference (4 total)45.5766.02593.01 / 616.97

At four scored candidates, gains are 0.68/1.50 AP points at similar latency, with paired 95% intervals [0.12,1.42]/[0.47,2.66]. Scoring 64 matches from the merged references closes the Real3D-AD gap, while AnyNormal retains a 1.34-point advantage on Anomaly-ShapeNet [0.21,2.59]. The 64-match method takes 2.93×/2.80× as long as scoring one match per reference, based on paired class-average scoring times.

We also varied neighborhood averaging. Without it, one match per reference versus four merged-reference matches gives 43.25/43.43 AP on Real3D-AD and 59.09/57.26 on Anomaly-ShapeNet. Selecting four merged-reference matches before averaging gives 46.10/64.33 AP across the two datasets. Thus reference choice and local context interact; the gains above apply to the scoring configurations in Table D2.

CPU timing includes test-cloud normals, neighborhood averaging, matching, scoring and pooling. Calls use one pinned core, single-threaded libraries and randomized method order on a shared server. Peak traced Python/NumPy allocations are 40.26 MiB for per-reference and 73.63 MiB for merged-reference scoring; whole-process peak resident memory is 883–885 MiB. Retrieved and scored match counts are controlled; internal KD-tree distance evaluations were not counted.

From an unregistered point cloud in host memory to a normal/anomaly decision, 30 strict-four-normal MulSen-AD objects take 22.92 s median / 41.44 s p95; scoring four merged-reference matches takes 22.85/41.43 s. Registration accounts for 97% of total per-reference time. These full-reference CPU timings include registration and normalization. The separate 20.46 ms RTX 4090 result uses compressed references and input already aligned to the category coordinate frame.

AuthorsOpenReview ↗

Response to Reviewer dXmL (2/3): support, registration and scorer training

dXmL — 2/3

Q3. Support size, diversity, missing modes and registration

Strict four-normal MulSen-AD. Refitting template, normalization, calibration and bank from the same four scans gives 85.51±1.96/81.67±0.64% object/point AUROC over all 15 categories and three support selections (Tables 1 and 16, pp. 5 and 20). Point AUROC exceeds the baseline results listed in Table 1(b), whose highest is 80.3% under its published protocol.

Normal-shape coverage. Values here follow Real3D-AD/Anomaly-ShapeNet/MulSen-AD. At fixed reference counts of 2/2/4, geometrically diverse supports outperform similar-shaped supports by 1.46/0.93/2.91 point-AUROC points after recalibration. Removing one cluster of geometrically similar normal scans gives 93.54/95.21/83.75%, compared with 95.06/96.24/87.04% using all references; removal changes both shape coverage and reference count.

With original coordinates and calibration fixed, increasing from two to four references raises point AUROC from 93.71 to 95.06% on Real3D-AD and 95.45 to 96.24% on Anomaly-ShapeNet. On MulSen-AD, four to all references raises it from 81.44 to 87.04%; Q5 gives the corresponding recall and threshold results.

Pose errors and recovery. We first add residual errors after alignment, without running registration again.

Table D3. Representative residual-alignment tests. Point AUROC (%); calibration fixed, query normals and neighborhoods recomputed. Translation is a fraction of the normal-reference bounding-box diagonal.

Residual errorReal3D-ADAnomaly-ShapeNetMulSen-AD
None95.0696.2487.04
Rotation 1°92.9093.4677.59
Translation 0.25%93.4894.2680.21

We separately reran registration on 12 predefined MulSen-AD objects: rotations of 5°/15°/45° about each axis, paired with translations of 2/5/10% of that diagonal, giving 108 nonzero conditions. Relative to registration without the added perturbation, 87/108 recover geometry within 0.5% of the diagonal and 107/108 within 1%; 88/108 meet rotation ≤2° and translation ≤0.5%.

Q4. Learned-scorer training and transfer

The 12→32→16→1 MLP uses Softplus hidden activations and, for the monotone control, positive effective weights. Its normal-only loss is

+0.25\operatorname{BCELogits}\left(\operatorname{TopMean}_{1\\%}\\{g_\theta(z):z\in O\\},0\right),$$ Here, $g_\theta$ outputs logits, and $P$/$O$ contain 4,096/8,192 sampled features from a normal scan. BCELogits is binary cross-entropy on logits; TopMean averages the largest 1% of logits. We use AdamW with learning rate **0.003**, weight decay **0.0001** and gradient-norm clipping **5**. We evaluate the **final scheduled checkpoint**, with no validation-based checkpoint selection (*Appendix E.8, Eq. 8 and Table 26, pp. 24–25*). The **toffees** model trained on four normal scans uses 30 epochs, training seed 20261210 and seed for ordering source categories 20261201. Real3D-AD/Anomaly-ShapeNet experiments varying the number of source categories use ordering seeds 20261201–20261203 and training seed 20261210; runs using all source categories use training seeds 20261101–20261103. The MulSen-AD experiment with four scans per source category uses 120 updates, category-ordering seeds 20261701–20261703 and training seeds 20261711–20261713. New unconstrained controls use four **airplane** scans, 120 updates and seeds 20261107–20261109. We use your suggested term, **scorer transfer with target-category normal support**: source-trained weights stay fixed, while target references supply shape information and calibration. The main fixed-mean scorer requires no training.
AuthorsOpenReview ↗

Response to Reviewer dXmL (3/3): reference-bank growth and quantile uncertainty

dXmL — 3/3

Q5. Bank expansion, false negatives and calibration trajectories

Adding references lowers scores for both normal and anomalous objects. AUROC improves as the bank grows, but keeping the four-reference threshold causes more misses.

Table D4. MulSen-AD bank expansion. Percentages averaged over three nested orders; original coordinates and calibration fixed. Recall is the maximum empirical ROC recall at FPR ≤5%. FNR/FPR retain the four-reference normal-test-score 95th-percentile threshold, used retrospectively. “All” uses all available normal training scans per category.

ReferencesObject / point AUROCROC recall@5%FPRFNR (fixed threshold)FPR (fixed threshold)
484.14 / 81.4459.3938.497.05
885.33 / 83.4060.9844.542.06
1686.36 / 84.9664.5147.860.75
3287.31 / 86.0668.0949.550.61
All88.96 / 87.0468.3250.520.55

The four-reference row differs from Q3 because it keeps the original coordinate system and calibration. Median normal/anomalous object-score ratios fall from 1/1 at four references to 0.25/0.89 at full support.

In a separate comparison, we re-estimate calibration at each reference count:

Table D5. Calibration as the MulSen-AD reference bank grows. Median ratios of the 99.5th-percentile residual scale to its full-support value, across category/order pairs; coordinates fixed.

ReferencesEuclideanPoint-to-planeNormal angle
41.212.622.99
81.091.782.10
161.071.461.55
321.031.221.28
All1.001.001.00

The geometric scales change with support size, especially for point-to-plane and normal-angle residuals. Table D4 holds these scales fixed to isolate the effect of adding references; Table D5 measures their change when calibration is re-estimated.

We traced nine selected cases that changed from detected to missed after adding a normal reference; eight had lower mean defect-region scores. In seven cases, changing only the background scores is enough to cross the old threshold (Table 22, p. 23). These traces show why the threshold needs to be checked when the reference bank changes.

Q6. Quantile estimation and four-normal uncertainty

Quantiles are estimated per category and residual type, shared across references and spatial locations. Leave-one-out comparisons of the four 16,384-point support scans give 65,536 finite residuals per residual type—approximately 328 values in the upper 0.5% of the sample. These residuals are spatially correlated, and the support still consists of four scans. We therefore assess uncertainty by resampling spatial blocks and leaving out scans.

We use 200 bootstrap replicates of spatial blocks from a 4×4×4 grid per scan. The intervals describe stability for these four observed scans with coordinates fixed.

Table D6. Four-normal point-to-plane quantile uncertainty. Category medians of 95% bootstrap interval width as a percentage of the original quantile estimate.

QuantileReal3D-ADAnomaly-ShapeNetMulSen-AD
95th28.928.4932.16
99.5th26.149.5425.42

Leaving out each scan in turn gives relative ranges of 34.4/12.6/42.1% for the 99.5th percentile, in the same dataset order.

We also test percentile pairs (90,99), (95,99), (95,99.5), (95,99.9) and (97.5,99.5). Using the main reference sets, point AUROC ranges are 94.91–95.20/96.12–96.34/86.71–87.42% in the same dataset order; MulSen-AD uses full support for this sweep. Separately, calibrating from four MulSen-AD scans with coordinates fixed gives 81.42%, versus 81.44% with full-bank calibration. This isolates calibration; Q3's strict-four-normal test recomputes every stage.

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

AnyNormal addresses few-shot 3D point-cloud anomaly detection, where a small set of normal scans may vary in local shape and sampling. It estimates normal geometric variation from normal-to-normal comparisons, calibrates Euclidean, point-to-plane, and normal-angle residuals at point and neighborhood levels, and scores each aligned normal reference separately before taking the pointwise minimum across references. The parameter-free scorer achieves 90.3/95.1 object/point AUROC on Real3D-AD, 94.5/96.2 on Anomaly-ShapeNet, and 89.0/87.0 on MulSen-AD. A learned monotone scorer transfers across datasets with little source-category data. A compressed configuration reaches up to 69.74 FPS on an RTX 4090, with a noticeable accuracy trade-off.

Critical strengths
  1. The method calibrates geometric discrepancies against observed normal variation and delays reference selection until each reference has been scored. Controlled comparisons report improvements over uncalibrated residuals and pooled matching, making the contribution and its source relatively clear.
  2. The paper reports results on three point-cloud benchmarks and tests cross-dataset scorer transfer. The fixed-mean variant avoids training a scoring network, while the compressed variant quantifies a practical speed-accuracy trade-off.
Critical weaknesses
  1. The method assumes aligned normal references and category-specific normal examples and calibration at test time. This is a meaningful operational requirement, so the results do not establish category-agnostic or reference-free detection.
  2. Compression reduces point AUROC considerably: for example, the reported Real3D-AD point AUROC drops from 95.1% to 82.4%, and MulSen-AD from 87.0% to 71.4%. The speed result is useful, but the paper should make the accuracy cost central when arguing for deployment efficiency. Comparisons with prior methods also retain their published preprocessing and protocols, making some cross-method differences less controlled than the component ablations.
Questions for authors

At deployment, which target-category information is required to build the aligned references and normal-variation calibration statistics, and are these computed exclusively from the allowed normal support set? This affects how the few-shot setting should be interpreted.

2 comments
AuthorsOpenReview ↗

Response to Reviewer SdDV: support protocol and accuracy–latency trade-off

Response to Reviewer SdDV

Thank you for highlighting the support requirements and the accuracy cost of fast inference. We clarify both below and provide a common-input comparison to complement the published-protocol results.

Q1. Target-category information and the four-normal protocol

Deployment requires the target category and its allowed normal support scans. We compute the alignment template, normalization, reference surfaces and normals, and normal-to-normal calibration statistics from these scans. Queries are aligned to this reference frame; test anomaly labels are used only for evaluation. The fixed-mean scorer needs no training. Scorer-transfer experiments reuse source-trained weights with each target category's own normal references and calibration.

For MulSen-AD, we evaluated a strict protocol in which the same four normal scans fit every component, including alignment and normalization. All 15 categories and 644 test objects were evaluated for each of three predetermined support selections, using one fixed registration profile across categories.

Table S1. MulSen-AD support comparison. Object/point AUROC (%); mean ± support-selection SD for strict four-normal support. Manuscript: Tables 1 and 16, pp. 5 and 20.

Normal supportObject AUROCPoint AUROC
Strict four-normal85.51 ± 1.9681.67 ± 0.64
Original full-normal88.9687.04

The strict four-normal point AUROC exceeds the published baseline results listed in Table 1(b), including Simple3D (80.3%) and MAD-3D (80.1%). Those baselines retain their published preprocessing protocols. The main table separates strict four-normal and full-support results; Real3D-AD and standard Anomaly-ShapeNet use four normal scans per category.

Accuracy–latency trade-off

The compressed configuration trades localization accuracy for low latency. Table S2 below compares both object detection and point localization.

Table S2. Full versus compressed accuracy. Object / point AUROC (%); standard 40-category Anomaly-ShapeNet.

DatasetFullCompressed
Real3D-AD90.30 / 95.0685.9 / 82.37
Anomaly-ShapeNet94.50 / 96.2490.1 / 88.00
MulSen-AD88.96 / 87.0479.3 / 71.39

Compression uses four 4,096-point references, point-to-plane scoring and sparse ICP; it also reduces MulSen-AD's support count.

On RTX 4090, median/p95 latency is 20.46/32.29 ms over 3,162 objects. Timing starts with a host-memory cloud in the category coordinate frame and includes fixed-index sampling, padding, transfers, sparse ICP, scoring, pooling and normal/anomaly decision. Coarse alignment and normalization take place before timing starts, and normal-reference preparation is done offline.

Comparison under common preprocessing

We also evaluated Simple3D on identical aligned 16,384-point inputs and four normal references, retaining its official pooling and method parameters (Section 3.2, p. 6).

Table S3. Common-input comparison. Object / point AUROC (%).

MethodReal3D-ADAnomaly-ShapeNet
Simple3D, common-input adaptation70.56 / 81.9283.31 / 89.67
AnyNormal, fixed mean90.30 / 95.0694.50 / 96.24

The advantage persists under shared inputs. We report this adaptation alongside the published-protocol results.

AuthorsOpenReview ↗

Gentle reminder

We would be grateful to hear whether our response on the four-normal support protocol and the accuracy cost of the compressed configuration addresses your concerns, and we are happy to clarify anything further before the discussion period closes.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the revision

We thank the reviewers for their careful reading. All changes are marked in the revised PDF.

Main message. A simple, training-free rule is enough for accurate few-shot 3D anomaly detection, and it generalizes. AnyNormal has no backbone, no learned scoring parameters and no synthetic anomalies. One fixed rule with the same constants is used for all 67 categories of three benchmarks. With four normal scans it gives the highest Real3D-AD object AUROC (90.3%) and Anomaly-ShapeNet point AUROC (96.2%) in Table 1(a), which includes networks trained with synthetic anomalies. Learned scorers reach the same accuracy and no more (Tables 4 and 26). A compressed configuration runs at 20.46 ms per object on an RTX 4090, with its accuracy cost reported beside it (Table 5).

The revision tests this message from each direction the reviewers raised. The table lists what was added and where to find it.

RequestAdded evidenceLocation
Four-shot MulSen-AD (SdDV Q, dXmL Q3, iQCc Q1)The same four scans fit alignment, normalization, calibration and references: 85.51±1.96 / 81.67±0.64 object/point AUROC, 15 categories × 3 support selectionsAbstract, Table 1(b), Sec. 3.2, Table 16
Calibration vs. neighborhood averaging (dXmL Q1)2×2 ablation on three datasetsTable 2(a)
Extra computation (dXmL Q2)Candidate budgets 1–64, CPU time, memory, full-pipeline latencyTables 2(b), 25, 27
Fixed constants (iQCc Q2, dXmL Q6)One-factor sweeps; four-scan quantile stabilityTables 17–18
Minimum rule and recall (dXmL Q5, iQCc Q4)Bank-size curves, recall at FPR ≤ 5%, threshold shift, per-reference tracesTables 19–20 and 22, Figure 6
Alignment and coverage (dXmL Q3, iQCc W2)Residual-error tests, registration recovery, spread vs. compact supportsTables 21, 23–24
Learned scorer (dXmL Q4, iQCc Q3)Objective, optimizer, seeds; unconstrained controlsEq. 8, Table 26
Accuracy cost of compression (SdDV W2)Compressed accuracy reported next to speedAbstract, Table 5
Shared preprocessing (SdDV W2)Simple3D on identical aligned inputs and four referencesSec. 3.2

What the new evidence shows

  1. The few-shot result holds on all three benchmarks. With four normal scans fitting every component, MulSen-AD point AUROC (81.67%) is above every listed baseline evaluated with the full normal training set (best: 80.3%).
  2. The gains come from the two named ideas. Calibration adds 5.48/1.35/4.10 point-AUROC points with neighborhood averaging held fixed. In point AP, every variant that scores candidates before selecting one outperforms selecting by proximity first (Table 25), and one score per reference matches or exceeds scoring 64 pooled candidates at about one third of the scoring time.
  3. A fixed rule is sufficient, and the controls explain why. Positive-weight scorers stay within 0.1 AUROC points of the fixed mean (Table 26). Without the positive-weight constraint, normal-only training reverses the ranking. The scorer needs the direction "larger calibrated deviation, higher score"; the fixed mean supplies it without training.
  4. The operating conditions are stated and measured. The method needs aligned inputs and target-category normal references, and the decision threshold should be set for the reference bank in use (Limitations, Appendix E). We describe the transfer result as scorer transfer with target-category normal support and do not claim category-agnostic or reference-free detection.
AuthorsOpenReview ↗

A note to the Area Chair and Reviewers on the contribution

Dear Area Chair and Reviewers,

Thank you for a thorough and constructive discussion. Your questions led to the most informative experiments in the revised paper. We would like to close by stepping back from the details to the main message.

Much recent work on 3D anomaly detection has moved toward larger systems: pretrained backbones, learned feature memories, reconstruction networks and synthetic anomalies. Our paper asks whether this complexity is necessary, and finds that it is not. Given four normal scans, it is enough to measure each geometric difference against how much normal objects already vary, and to let each region choose the normal example that explains it best. One training-free rule then performs at the level of trained detectors across three benchmarks and 67 categories, and a compressed version runs in tens of milliseconds per object.

We believe this result is useful to the field in three ways. It provides a strong, simple baseline against which more complex methods can be measured. It shows which ingredients carry the signal, since each is isolated by a controlled experiment. And because the method is small and inspectable, it is easy to reproduce and apply in practice.

The reviewers recognized these strengths in their own words: a contribution whose source is "relatively clear" (SdDV), a simplicity that makes it "easy to inspect and potentially useful in practice" (dXmL), and "a well-supported negative result" (iQCc). During the rebuttal we tested the main concerns directly: the four-shot setting, the source of the gains, computation, robustness and the learned scorer. The central finding held in each case.

We hope the reviewers will take this evidence into account in their final assessment. Simple ideas that are carefully tested are easy to overlook, and we believe this one deserves a place in the discussion of how few-shot 3D anomaly detection should be done.

Thank you again for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.