acceptodds
Under review as a conference paper at ICLR 2027

Training-Free 3D Anomaly Detection via Viewpoint-Preserving Alignment and Residual Calibration

Abstract

3D anomaly detection aims to identify defective objects and localize structural deviations from a small set of normal observations. Existing methods often detect anomalies indirectly through learned or engineered representations, or through discrepancies from reconstructed normal geometry. For structural anomalies, however, the anomaly signal is inherently geometric, suggesting a simpler and direct alternative: align a query point cloud with normal references and directly measure their geometric residuals. The challenge is that large residuals are not unique to defects; they can also arise from alignment errors, partial observations, and sampling noise. In this work, we argue that geometric residuals become meaningful anomaly cues when pose-induced discrepancies are controlled and residual evidence is evaluated consistently across spatial scales. Motivated by this principle, we introduce MARS, a training-free 3D anomaly-detection method that addresses the residual ambiguity issues through viewpoint-preserving Multi-hypothesis Alignment and normal-calibrated Residual Scoring. MARS first generates candidate poses from precomputed reference views and preserves alternative viewing directions during hypothesis screening before selecting the final pose. Then, query-to-normal residuals are summarized across multiple spatial scales and evaluated against the corresponding residual statistics observed when normal references are compared with one another. With only a few normal references per category and no pretrained models, training, or synthesized anomalies, MARS achieves 97.0%/91.2% point-/object-level AUROC on Real3D-AD and 98.2%/95.7% on Anomaly-ShapeNet, outperforming recent methods on average.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer vision, Anomaly Detection, 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_v4fEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MARS, a training-free method for 3D anomaly detection using a few normal point clouds. The method combines viewpoint-preserving multi-hypothesis alignment with normal-calibrated residual scoring. It first aligns queries with normal references, then evaluates multi-scale geometric residuals against normal-reference statistics. Experiments on Real3D-AD and Anomaly-ShapeNet achieve point-/object-level AUROC scores of 97.0%/91.2% and 98.2%/95.7%, respectively. Ablations and efficiency analyses further support the proposed design.

Comment

This paper presents a simple and effective training-free approach with strong results across two 3D anomaly detection benchmarks. The experimental analysis is generally convincing. While the independent benefit of normal calibration and the robustness to reference selection and missing structures require further clarification, the overall contribution is meaningful, and I lean toward acceptance.

Critical strengths

1. Simple and effective approach. MARS performs anomaly detection directly from geometric differences without pretrained models or training. The combination of multi-hypothesis alignment and normal-reference comparison is practical, particularly when only a few normal samples are available.

2. Strong experimental results. The method outperforms the reported baselines on average across both benchmarks. The alignment and residual-scoring ablations, robustness tests, and runtime comparisons provide useful evidence of its effectiveness.

Critical weaknesses

1. The benefit of normal calibration is not fully established.

Although normal-calibrated residual scoring is a central contribution, Appendix Table D.4 shows that uncalibrated aggregation already achieves comparable performance. On Anomaly-ShapeNet, the uncalibrated mean even slightly outperforms calibrated Fisher-style scoring in object-level AUROC (95.93% vs. 95.66%). A more controlled comparison would help clarify the actual contribution of normal calibration beyond multi-scale residual aggregation.

2. Limited validation of reference selection and missing-structure anomalies.

MARS relies on a small set of normal references, but the reference-count experiment uses only one deterministic two-reference subset. Performance variation across different reference selections remains unclear. Moreover, the directed query-to-reference residual cannot directly detect missing regions that contain no observed points. This may limit the method's applicability to certain industrial defects.

Questions for authors

Q1. Can the authors further isolate the contribution of normal calibration from multi-scale residual aggregation? Appendix Table D.4 suggests that calibration provides only modest improvements over uncalibrated scoring.

Q2. How sensitive is MARS to the selection of normal references? Could the authors report results across multiple randomly selected two-shot and four-shot reference sets?

Q3. How does MARS perform specifically on missing-part or missing-material defects? Can the authors provide a defect-type analysis to clarify the limitations of directed geometric residuals?

4 comments
AuthorsOpenReview ↗

Response to the Reviewer v4fE [1/3]

Dear Reviewer v4fE, the authors sincerely appreciate your thoughtful review and positive assessment. Your questions helped us isolate the contribution of normal calibration and state the limits of a directed residual precisely. We are also thankful for your acknowledgement of the strengths of this work: a simple and effective training-free approach that detects anomalies directly from geometric differences, supported by ablations, robustness tests and runtime comparisons. We address your concerns and questions below.

W1/Q1) Isolating normal calibration. ... On Anomaly-ShapeNet, the uncalibrated mean even slightly outperforms calibrated Fisher-style scoring in object-level AUROC (95.93% vs. 95.66%). A more controlled comparison would help clarify the actual contribution of normal calibration beyond multi-scale residual aggregation. ... Can the authors further isolate the contribution of normal calibration from multi-scale residual aggregation?

Answer: The row “Uncalibrated mean” averages the seven raw statistics without normal-bank tail calibration; we renamed it “Mean” under “Anomaly scoring from raw residual statistics” (Table D.7(b), formerly Table D.4(b)). With the seven multi-scale statistics and Fisher-style fusion fixed, normal LOO calibration improves both P- and O-AUROC on both benchmarks (Table R13).

Table R13. Four-shot comparison of per-query and normal leave-one-out (LOO) tails with Fisher-style fusion (P/O-AUROC, %). Chamfer refinement and median smoothing are disabled.

BenchmarkPer-query ranks + FisherNormal LOO + FisherGain
Real3D-AD95.78 / 83.1996.88 / 90.73+1.10 / +7.54
Anomaly-ShapeNet97.64 / 94.5098.17 / 95.60+0.53 / +1.10
  • Normal-reference tails improve P/O-AUROC by 1.10/7.54 pp on Real3D-AD and 0.53/1.10 pp on Anomaly-ShapeNet. The gain comes from ranking the same multi-scale residuals against the normal reference distribution instead of each query’s own distribution.
  • At two shots, the gains averaged over all six reference pairs are 0.77/8.86 pp on Real3D-AD and 0.49/1.67 pp on Anomaly-ShapeNet.

Table R13 disables post-processing because Chamfer refinement uses normal-scan thresholds. With post-processing enabled and Fisher-style fusion fixed, normal LOO improves four-shot P/O-AUROC over per-query tails by 1.01/2.99 pp on Real3D-AD and 0.45/1.06 pp on Anomaly-ShapeNet.

As you note, directly averaging the raw residual statistics already comes close to our final scorer on Anomaly-ShapeNet (98.16/95.93% vs. 98.19/95.66% P/O-AUROC, Table D.7(b–c)), which supports our main finding that aligned multi-scale residuals carry most of the signal. The plain average also drops Fisher-style fusion, so it cannot isolate calibration; Table R13 does. Against the plain average, the gain from calibration is small but consistent, and it is in localization: the full scorer localizes better in all 12 Real3D-AD and 31 of 40 Anomaly-ShapeNet categories (sign test, p < 0.001 on each benchmark), although the average gain is only 0.76 and 0.03 pp. Calibration puts the seven statistics, whose normal ranges differ, on one scale so that Fisher-style fusion can let the most unusual one dominate; averaging the calibrated scores instead falls below the plain average (94.73 vs. 96.19 P-AUROC on Real3D-AD, Table D.7(c)). For detection, raw aggregation does as well, so we claim no detection gain over it (Sec. 4.3.2, Lines 515–517).

We thank the reviewer; these comparisons clarify what normal calibration adds. With Fisher-style fusion fixed, its largest gain is in object-level detection on partial real scans (Table R13); against the plain average, its gain is small but consistent and in localization. They are in Sec. 4.3.2 (Table 6, Lines 495–502 and 513–517) and App. D.7 (Table D.7, Lines 1188–1213 and 1224–1236).

AuthorsOpenReview ↗

Response to the Reviewer v4fE [2/3]

W2-1/Q2) Choice of normal references. MARS relies on a small set of normal references, but the reference-count experiment uses only one deterministic two-reference subset. Performance variation across different reference selections remains unclear. ... How sensitive is MARS to the selection of normal references? Could the authors report results across multiple randomly selected two-shot and four-shot reference sets?

Answer: MARS is robust to the choice of references: across all six two-reference pairs, the standard deviation of P/O-AUROC is at most 0.70 pp, and even the worst pair beats the strongest two-shot baseline. For each pair, we rebuilt the whole pipeline and evaluated all 2,518 queries.

Table R14. Reference-subset sensitivity (P / O-AUROC, %; two-shot mean ± SD and range over six subsets).

DatasetTwo-shot PTwo-shot OFour-shot P / O
Real3D-AD95.73 ± 0.70 (94.41–96.31)88.50 ± 0.70 (87.47–89.22)96.95 / 91.23
Anomaly-ShapeNet98.03 ± 0.07 (97.91–98.10)94.98 ± 0.49 (94.10–95.53)98.19 / 95.66
  • Partial scans are more sensitive to the choice: the ranges are 1.90/1.75 pp on Real3D-AD and 0.19/1.43 pp on Anomaly-ShapeNet.
  • Only one four-reference set exists per category: both benchmarks provide exactly four normal training scans [G1, G2], so independent four-shot sets cannot be drawn; the six pairs cover every two-reference choice.
  • Even the lowest P- and O-AUROC across the six MARS pairs exceed the two-shot Reg2Inv results reported in Table 1 (90.6/80.2 on Real3D-AD and 90.3/84.5 on Anomaly-ShapeNet). Reg2Inv was evaluated on one reference pair.

We thank the reviewer; this analysis shows that MARS works with fewer, differently chosen references. It is in Sec. 4.2.1 (Lines 371–375) and App. E.1 (Lines 1341–1345; Table E.1(c), Lines 1368–1373).

AuthorsOpenReview ↗

Response to the Reviewer v4fE [3/3]

W2-2/Q3) Missing-part and missing-material defects. Moreover, the directed query-to-reference residual cannot directly detect missing regions that contain no observed points. This may limit the method's applicability to certain industrial defects. ... How does MARS perform specifically on missing-part or missing-material defects? Can the authors provide a defect-type analysis to clarify the limitations of directed geometric residuals?

Answer1: MARS and the baselines generally score lower on broken, hole, and crack defects than on bulges and concavities. These queries have fewer labeled defect points, which more often form two closely spaced layers.

Table R15. Results by defect type (category-averaged AUROC, %); the baseline columns report the highest P- and O-AUROC among four rerun methods.

DatasetDefect typeAnomalous queriesMARS P / OBest baseline P (method)Best baseline O (method)
Anomaly-ShapeNetBulge28099.57 / 100.0095.41 (Reg2Inv)95.33 (PASDF)
Concavity28099.68 / 99.7695.66 (Reg2Inv)93.07 (PASDF)
Broken3668.32 / 60.9371.94 (PointCore)61.30 (PointCore)
Hole3864.56 / 63.3367.68 (Reg2Inv)57.54 (PASDF)
Crack2562.85 / 73.8561.48 (PASDF)62.05 (PASDF)
Real3D-ADBulge42697.55 / 92.5891.55 (Reg2Inv)82.54 (Reg2Inv)
Sink16492.14 / 85.1884.66 (Reg2Inv)77.48 (PASDF)
Hybrid1294.10 / 97.3381.85 (Reg2Inv)91.43 (PointCore)
  • MARS trails the best baseline only on Broken and on Hole localization. PointCore is better on Broken in both metrics and Reg2Inv on Hole in P-AUROC; MARS leads in every other column.
  • These defects leave few labeled points, often in two thin layers. Broken, Hole, and Crack queries have about one-seventh as many labeled defect points (36.72 vs. 257.62), and 81.8% of them form two closely spaced layers (vs 34.8%; Table R16). Defect points that stay close to the normal surface after alignment have small residuals, so few of them stand out among top-scoring points.

Table R16. Labeled defect point counts and spatial distribution on Anomaly-ShapeNet.

MeasureBroken, Hole, Crack (99 queries)Bulge, Concavity (560 queries)
Mean GT defect points per query36.72257.62
Queries with a two-layer pattern¹81/99 (81.8%)195/560 (34.8%)
Mean layer gap (% of query diagonal)1.362.04

¹ The labeled defect points form two groups at different depths in the same local region.

Answer2: A region with no observed points cannot be scored directly, so MARS detects a missing part only through the surface around it. It flags the object when an exposed cut surface or broken edge deviates from the normal scans, and when removed material leaves a scanned floor, as for Real3D-AD sinks, MARS reaches 92.14/85.18 P/O-AUROC, above every baseline. Neither benchmark has a separate missing-part category; Anomaly-ShapeNet’s Hole and Broken types are the closest. Sec. 5 (Lines 521–526) states this limitation and a remedy, a visibility-aware comparison from the normal scans to the query.

We thank the reviewer; this analysis clarifies how MARS performs across defect types and why missing regions with no observed points are difficult to detect. The defect-type results are in App. E.4 (Tables E.3–E.4, Lines 1458–1469 and 1495–1520).

We hope our responses address your concerns. If any question remains, please let us know, and we will respond promptly.

I appreciate the author’s detailed rebuttal, which has largely addressed my concerns. I have decided to maintain my original score and recommend acceptance.

Reviewer_6tbYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MARS, a training-free framework for 3D point cloud anomaly detection. It aligns query point clouds with normal references through viewpoint-preserving multi-hypothesis registration, then combines multi-scale neighborhood means and standard deviations of geometric residuals using leave-one-out normal calibration. Experiments on Real3D-AD and Anomaly-ShapeNet demonstrate strong detection and localization performance using four normal references per category, without pretrained models or synthetic anomalies.

Critical strengths
  1. Multi-scale neighborhood means and standard deviations turn pointwise geometric residuals into more effective evidence for defect localization. This insightful design yields consistent, substantial localization improvements on both datasets.
  2. The method requires only a few normal references, without model training, pretrained representations, or synthetic anomalies. Its strong performance on both real partial scans and complete synthetic shapes demonstrates practical value.
Critical weaknesses
  1. The independent benefit of normal-reference calibration remains limited, as simple uncalibrated multi-scale aggregation already approaches the full method’s performance. A comparison against the best single-scale variant is also missing, leaving the respective contributions of neighborhood aggregation and multi-scale complementarity unclear.
  2. The method depends on the selected pose, but lacks a clear analysis attributing detection failures to either registration errors or the scoring stage’s inability to distinguish normal variation from defects after alignment.
Questions for authors
  1. Given the small gains over uncalibrated averaging in Table D.4(b), which categories or conditions benefit most from normal-reference calibration, and how consistent are these gains?
  2. With registration and post-processing fixed, how does multi-scale aggregation compare with the best single-scale variant?
  3. Can you provide representative failure cases distinguishing registration failures from scoring failures, and clarify whether calibration improves robustness to residual pose errors?
4 comments
AuthorsOpenReview ↗

Response to the Reviewer 6tbY [1/3]

Dear Reviewer 6tbY, The authors sincerely appreciate your constructive review and the precise questions on normal calibration, the neighborhood scales and the failure cases. They helped us show what each scoring component contributes and where MARS can fail. We are also thankful for your acknowledgement of the strengths of this work: insightful multi-scale residual statistics that consistently improve localization, and practical training-free detection from only a few normal references. We address your concerns and questions below.

W1-1/Q1) What normal calibration adds. The independent benefit of normal-reference calibration remains limited, as simple uncalibrated multi-scale aggregation already approaches the full method’s performance. ... Given the small gains over uncalibrated averaging in Table D.4(b), which categories or conditions benefit most from normal-reference calibration, and how consistent are these gains?

Answer1: The plain average of the seven raw statistics already performs well (96.19/90.96 on Real3D-AD, 98.16/95.93 on Anomaly-ShapeNet), which supports our main finding that aligned multi-scale residuals carry most of the signal. This row, formerly "Uncalibrated mean", uses no tail probabilities, so we renamed it “Mean” under “Anomaly scoring from raw residual statistics” (Table D.7(b), formerly Table D.4(b)). Against it, the full scorer gains 0.76/0.27 pp P/O-AUROC on Real3D-AD and 0.03 pp P-AUROC on Anomaly-ShapeNet, where the plain average stays 0.27 pp higher in O-AUROC.

Answer2: Normal calibration pays off together with Fisher-style fusion: ranking each statistic against the normal scans instead of within the query raises P/O-AUROC by 1.10/7.54 pp on Real3D-AD and 0.53/1.10 pp on Anomaly-ShapeNet. Fisher-style needs each statistic as a tail probability, which can come from the query itself or from the normal scans. Averaging the normal-bank scores (1 − p) yields lower P-AUROC than averaging the raw statistics (94.73 vs. 96.19 on Real3D-AD). Fisher-style fusion gives greater weight to small tail probabilities, so an unusually large residual in one channel can contribute more to the anomaly score (Table D.7(b)–(c)).

Table R10. Per-query vs. leave-one-out (LOO) normal-reference tails with Fisher-style fusion (P / O-AUROC, %).

BenchmarkShotsPer-query ranks + FisherNormal LOO + FisherGain
Real3D-AD495.78 / 83.1996.88 / 90.73+1.10 / + 7.54
294.90 / 79.0095.67 / 87.86+0.77 / +8.86
Anomaly-ShapeNet497.64 / 94.5098.17 / 95.60+0.53 / + 1.10
297.52 / 93.2498.01 / 94.91+0.49 / +1.67
  • We disable post-processing because Chamfer refinement uses normal-scan thresholds. With post-processing enabled, four-shot P/O-AUROC gains are +1.01/+2.99 pp on Real3D-AD and +0.45/+1.06 pp on Anomaly-ShapeNet.
  • Mean over the six pairs; the gain is positive for both metrics in every pair.

Answer3: The gains are largest for object-level detection on the partial real scans, with fewer references and for the most common defect type, and localization improves in most categories.

  • Detection on partial scans gains most, and more with fewer references. Real3D-AD O-AUROC gains the most (Table R10); with two references the gain grows to 8.86 pp, and for bulges, the most common defect (426 out of 602 anomalous Real3D-AD scans), it is 1.22/9.83 pp.
  • Localization improves in most categories. P-AUROC improves in 10 of 12 Real3D-AD and 36 out of 40 Anomaly-ShapeNet categories, most for diamond and ashtray0; O-AUROC improves in 11 out of 12 Real3D-AD categories, while on Anomaly-ShapeNet, where MARS already reaches 100% in 21 of 40 categories (Table C.2), it ties in 17 (Table R11).
  • The full scorer also localizes better than the plain average in every Real3D-AD category and in 31 of 40 Anomaly-ShapeNet categories (sign test, p < 0.001 on each benchmark).

Table R11. Categories whose AUROC improves / ties / declines when ranking against the normal scans (conducted under the same condition as R10).

BenchmarkShotsP-AUROCO-AUROCLargest ΔPLargest ΔO
Real3D-AD410 / 0 / 211 / 0 / 1+3.30 (diamond)+28.96 (diamond)
210 / 0 / 211 / 0 / 1+2.71 (diamond)+31.71 (diamond)
Anomaly-ShapeNet436 / 0 / 414 / 17 / 9+4.60 (ashtray0)+23.49 (bucket0)
233 / 0 / 726 / 11 / 3+4.37 (ashtray0)+21.01 (bucket0)

We thank the reviewer; this comparison shows that normal LOO calibration improves both localization and detection on both benchmarks with Fisher-style fusion fixed. It is in Sec. 4.3.2 (Lines 513–517; Table 6, Lines 495–502) and App. D.8 (Table D.8, Lines 1238–1257).

AuthorsOpenReview ↗

Response to the Reviewer 6tbY [2/3]

W1-2/Q2) Multi-scale vs. the best single scale. A comparison against the best single-scale variant is also missing, leaving the respective contributions of neighborhood aggregation and multi-scale complementarity unclear. ... With registration and post-processing fixed, how does multi-scale aggregation compare with the best single-scale variant?

Answer: No single scale is best for both localization and detection, while all four scales together come within 0.68 pp of the best of each without choosing a scale per metric. MARS summarizes residuals over four neighborhood sizes k (1, 16, 64, 256): the mean at each size and standard deviation at the three larger ones. For each size, we test the mean alone and the mean with the SD, keeping registration, calibration, Fisher-style fusion and post-processing fixed (Table R12).

Table R12. Single scales vs. all four scales (P / O-AUROC, %). Bold indicates the best value in each dataset and metric.

Residual StatisticsReal3D-ADAnomaly-ShapeNet
Mean, k = 183.66 / 91.8590.36 / 95.46
Mean, k = 1688.01 / 91.0494.50 / 94.61
Mean, k = 6494.82 / 87.8996.27 / 91.80
Mean, k = 25695.93 / 83.7095.60 / 87.07
Mean + std, k = 1688.75 / 91.9195.35 / 95.37
Mean + std, k = 6496.14 / 91.2297.43 / 95.20
Mean + std, k = 25696.96 / 87.7797.02 / 94.16
All 4 scales, 7 statistics (MARS)96.95 / 91.2398.19 / 95.66
  • On Real3D-AD, the best single scale differs by metric: k = 256 (mean+SD) localizes best and k = 16 (mean+SD) detects best, and all four scales together come within 0.01 and 0.68 pp of them.
  • On Anomaly-ShapeNet, all four scales together beat every single scale on both metrics.
  • The standard deviations add information the means lack: adding the three SD statistics to the four means raises P/O-AUROC from 95.52/89.44 to 96.95/91.23 on Real3D-AD and from 97.51/95.07 to 98.19/95.66 on Anomaly-ShapeNet.

We thank the reviewer; this comparison shows that combining scales spares MARS a per-dataset, per-metric choice of neighborhood size. It is in App. D.4 (Lines 1072–1079 and 1109–1110; Table D.4, Lines 1090–1106).

AuthorsOpenReview ↗

Response to the Reviewer 6tbY [3/3]

W2/Q3) Registration failures vs. scoring failures. The method depends on the selected pose, but lacks a clear analysis attributing detection failures to either registration errors or the scoring stage’s inability to distinguish normal variation from defects after alignment. ... Can you provide representative failure cases distinguishing registration failures from scoring failures, and clarify whether calibration improves robustness to residual pose errors?

Answer1: These cases illustrate two possible contributors to failure: pose error, which a better pose can largely reduce, and normal shape variation, which can persist after improved alignment.

  • A wrong pose (duck, Fig. E.1): a better pose removes most of the mismatch. For a normal duck scan, the selected pose leaves a visible mismatch; applying the saved pose of a corresponding scan lowers the median query-to-reference distance from 0.928 to 0.097. Its object score stays high (140.74 vs. 145.56), however, so the false alarm is not due to the pose alone.
  • Normal variation (airplane, Fig. E.2): the high scores remain after a better pose. The body aligns closely, but high scores concentrate on the tail, whose shape differs even among the normal references. Correcting the pose on the body does not remove them. Without ground-truth poses, we cannot rule out a small remaining pose effect.

Answer2: Calibration is not designed to correct pose errors; MARS’s robustness to the pose comes from re-registering every query. A random rotation and a translation of up to 10% of the query’s bounding-box diagonal along each axis, applied before registration, change P/O-AUROC by at most 0.75 pp (Table E.2). Since neither benchmark provides ground-truth poses, we cannot measure how much of the calibration gain comes from tolerating small remaining pose errors; it is largest on the partial real scans of Real3D-AD (+7.54 pp O-AUROC; Table R10), where partial-to-full registration can be more challenging than full-to-full registration.

We thank the reviewer; examining these cases helps clarify how pose mismatch and normal shape variation can contribute to false alarms, even when their effects cannot be fully disentangled. The cases are in App. E.3 (Lines 1390–1402 and 1427–1430; Fig. E.1, Lines 1449–1452; Fig. E.2, Lines 1490–1493), and registration residuals by ground-truth region in Sec. 4.3.1 (Table 3, Lines 457–468 and 475–480).

We hope our responses clarify how calibration, the neighborhood scales and registration each contribute. If any question remains, please let us know, and we will respond promptly.

I thank the authors for taking the time to respond to my comments. After reading all the reviews and responses, I maintain my positive score.

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

This paper presents MARS, a training‑free few‑reference 3D structural anomaly detector. It leverages query‑normal geometric residuals as anomaly signals by mitigating pose artifacts and calibrating residuals against normal‑sample variations. MARS conducts viewpoint‑preserving multi‑hypothesis alignment: normal references are aligned to a canonical anchor; pose hypotheses are generated via 42‑view depth rendering and polar‑Fourier roll estimation. One candidate per viewpoint survives for ICP refinement, and the final pose is picked using lower‑median residuals on shared query points. After alignment, median‑normalized residuals against pooled references form multi‑scale mean‑std channels. Leave‑one‑out normal comparisons produce empirical distributions; upper‑tail probabilities are Fisher‑aggregated, followed by Chamfer‑based false‑positive suppression and median smoothing. Object‑level scores use harmonic mean of top‑16 point scores. With four references, MARS achieves 97.0%/91.2% point‑/object‑AUROC on Real3D‑AD and 98.2%/95.7% on Anomaly‑ShapeNet, retaining strong results with two references. Ablations verify critical roles for ICP and multi‑hypothesis alignment, as well as gains from multi‑scale statistics and normal calibration; analyses also cover registration quality and robustness to sparsity and noise.

Critical strengths
  1. MARS needs neither a pretrained network, anomalous examples, nor synthesized defects, yet performs strongly on both a partial real-scan benchmark and a complete synthetic-shape benchmark.
  2. The paper separately studies alignment, roll estimation, pruning, residual channels, calibration, reference merging, sparsity, noise, and two-shot behavior. In particular, the large degradation without ICP and the progressive gains from multiscale statistics and calibration make the source of performance more credible.
Critical weaknesses
  1. Multi-view pose hypotheses, polar/Fourier correlation, ICP refinement, nearest-neighbor geometric residuals, local means and variances, empirical tail calibration, Fisher aggregation, Chamfer-based attenuation, median filtering, and top-k object aggregation are all established ideas. The contribution is primarily a carefully engineered sequence and robust selection policy rather than a new representation, objective, or inference principle.
  2. The many constants and aggregation rules create substantial benchmark-tuning capacity without a stated independent selection protocol. Most four-shot baselines are imported under potentially different reference, sampling, and aggregation protocols, and reported query time excludes reference registration and calibration. Thus, it is unclear whether the gains arise from a general principle or dataset-specific engineering and evaluation choices.
Questions for authors
  1. On what data were all alignment, calibration, attenuation, smoothing, and top‑16 aggregation hyperparameters chosen? Could the authors report results when every hyperparameter is fixed on one dataset and transferred without adjustment to the other?
  2. Could the authors rerun the strongest feasible baselines using exactly the same four references, 8,192 query points, input preprocessing, and object‑score aggregation? For results imported from external papers, could they enumerate protocol differences that may influence each comparison?
  3. I would like to see category setup time, reference‑alignment time, calibration time, peak CPU/GPU memory, and overall end‑to‑end latency alongside online query‑processing runtime. How many ICP runs happen on average under pruned conditions, rather than under worst‑case scenarios?
  4. The two‑shot experiment picks the lexicographically first two references. Could the authors provide mean and variance computed across multiple or all two‑reference subsets, and discuss how sensitive performance is with respect to reference diversity?
  5. Which components are truly essential when compared against a strong multi‑hypothesis registration baseline? Could the authors conduct incremental ablations starting from best‑possible alignment plus a single pre‑registered robust residual score, then add calibration, Fisher aggregation, attenuation, and smoothing step‑by‑step, while reporting confidence intervals across both datasets?
  6. Empirical tail probabilities are estimated from a very small number of leave‑one‑reference‑out samples, and Fisher aggregation typically builds upon assumptions of well‑calibrated, weakly‑dependent p‑values. I am curious how ranking outputs respond to channel discreteness and inter‑channel dependence. Could the authors also include comparisons against simple normal‑only empirical‑rank aggregation and learning‑free Mahalanobis aggregation?
7 comments
AuthorsOpenReview ↗

Response to the Reviewer ACHj [1/6]

Dear Reviewer ACHj, the authors sincerely appreciate your careful review and your concrete questions on the evaluation protocol, the hyperparameters and the source of MARS's gains. They helped us identify which components matter, compare all methods under one protocol, and state our contribution more precisely. We are also thankful for your acknowledgement of the strengths of this work: training-free detection without pretrained networks, anomalous examples or synthesized defects, strong performance on both partial real scans and complete synthetic shapes, and ablations that make the source of performance credible. We address your concerns and questions below.

W1) Contribution beyond established components. Multi-view pose hypotheses, polar/Fourier correlation, ICP refinement, nearest-neighbor geometric residuals, ... are all established ideas. The contribution is primarily a carefully engineered sequence and robust selection policy rather than a new representation, objective, or inference principle.

Answer: We agree that each tool in MARS is classical. Our contribution is the finding that a basic signal, the distance from each aligned query point to the nearest normal point, can support strong anomaly detection across partial real scans and complete synthetic shapes. Recent 3D anomaly detectors mainly improve the representation by learning features [G3, G5] or reconstructing normal shapes [G2]. We instead return to this distance (the residual) and identify what makes it reliable.

  • The closest earlier detectors either skip the alignment or do not rely on the aligned residual alone. BTF [G6] showed that classical descriptors can rival learned ones. It matches features against a memory bank without aligning the query to the normal scans and achieves 63.5% O-AUROC on Real3D-AD (Table 1). Other registration-based methods use learned features alongside registration [G3, G5]. Reg3D-AD [G1] combines raw-point and learned-feature scores, and MVR-DR [G9] combines geometric and feature-space residuals.
  • Condition 1: the pose must be right even for a partial scan. MARS keeps the best candidate from each viewing direction for final refinement. Removing viewpoint search lowers P/O-AUROC from 96.95/91.23% to 63.16/58.58% for partial-to-full registration on Real3D-AD and from 98.19/95.66% to 72.23/73.81% for full-to-full registration on Anomaly-ShapeNet (Tables 4 and D.1(a)).
  • Condition 2: summarize residuals across neighborhood sizes. Multi-scale statistics raise P-AUROC by 15.49 pp on Real3D-AD in the cumulative scoring ablation (Table D.2). With the same seven channels and Fisher-style fusion fixed, using the normal LOO distribution in place of the per-query distribution further improves O-AUROC by 7.54 pp before post-processing (Table 6).
  • With both conditions met, classical tools are enough to outperform trained detectors. MARS exceeds the strongest competitor on each Real3D-AD metric by 8.8 pp P-AUROC (97.0 vs. 88.2, CASL) and 8.3 pp O-AUROC (91.2 vs. 82.9, PointCore) without any training, so the gain comes from using the right signal under the right conditions rather than from a new learned component.

We thank the reviewer for this question, which helped us state the finding and its relation to prior work more clearly. Sec. 1 (lines 89–107) states our finding, and Sec. 2 (lines 123–138) discusses related registration-based methods.

AuthorsOpenReview ↗

Response to the Reviewer ACHj [2/6]

W2-1/Q1) Hyperparameter selection. The many constants and aggregation rules create substantial benchmark-tuning capacity without a stated independent selection protocol. ... On what data were all alignment, calibration, attenuation, smoothing, and top-16 aggregation hyperparameters chosen? Could the authors report results when every hyperparameter is fixed on one dataset and transferred without adjustment to the other?

Answer1: As in the submission, MARS uses one configuration for all 52 categories of both benchmarks; because neither benchmark has a validation split, we chose it using Real3D-AD test results. Both benchmarks provide only four normal training scans per category and a test set [G1, G2]. Prior work has made specific choices the same way, such as the RIConv++ neighborhood size in Reg2Inv (App. B.2 of [G3]) and the number of patches for pseudo-anomalies in PO3AD (Sec. 4.6 of [G8]).

Answer2: Performance changes little across the tested hyperparameter settings. Varying one setting at a time changes P/O-AUROC by at most 0.23/1.14 pp on Real3D-AD and 0.07/0.52 pp on Anomaly-ShapeNet (Table R1). We use 42 viewpoints because increasing to 92 improves P/O-AUROC by only 0.08/0.06 pp while raising inference time from 1.04 to 1.65 s per query (Fig. 4(a–c)), so we did not simply take the setting with the best test score. The pooling and point-count results are also reported in Tables D.9(b) and E.1(a).

Table R1. One setting varied at a time (P/O-AUROC, %).

SettingValueReal3D-ADAnomaly-ShapeNet
Viewpoints per reference3296.80 / 90.7098.16 / 95.23
42 (default)96.95 / 91.2398.19 / 95.66
9297.03 / 91.2998.23 / 95.31
Object pooling¹Max96.95 / 91.0498.19 / 95.63
Top-16 arithmetic mean96.95 / 91.1898.19 / 95.66
Top-16 harmonic mean (default)96.95 / 91.2398.19 / 95.66
Scored points4,09697.14 / 90.0998.21 / 95.14
8,192 (default)96.95 / 91.2398.19 / 95.66

¹ Pooling uses the same point scores, so P-AUROC does not change.

The gain over the baselines does not depend on these settings. With the same alignment, a single residual statistic ranked against the normal scans already reaches 91.1% O-AUROC on Real3D-AD, above every baseline in Table 1 (at most 82.9%); the multi-scale statistics and fusion mainly add localization (P-AUROC from 79.7% to 97.0%, Table 5).

Answer3: Because the configuration was chosen on Real3D-AD, our Anomaly-ShapeNet results are already a transfer result, and they are within 0.01 pp of the configuration chosen on Anomaly-ShapeNet itself. Selecting every setting on Anomaly-ShapeNet changes only one choice, turning Chamfer refinement off; transferred back to Real3D-AD, this costs 0.42 pp O-AUROC (Table R2, from the “No Chamfer refinement” row of Table D.9(a)).

Table R2. Cross-benchmark transfer (P/O-AUROC, %). When settings were selected separately on each benchmark, only Chamfer refinement differed among the settings tested.

Configuration selected onChamfer refinementEvaluated on Real3D-ADEvaluated on Anomaly-ShapeNet
Real3D-AD (paper default)On96.95 / 91.2398.19 / 95.66 (transfer)
Anomaly-ShapeNetOff96.95 / 90.81 (transfer)98.19 / 95.67

We thank the reviewer for this question; these analyses show that one configuration of MARS works across benchmarks and that its results are insensitive to the settings. We state the selection protocol in Sec. 4.1 (Lines 350–353) of the revised PDF.

AuthorsOpenReview ↗

Response to the Reviewer ACHj [3/6]

W2-2/Q2) Baselines under the same protocol. Most four-shot baselines are imported under potentially different reference, sampling, and aggregation protocols ... Could the authors rerun the strongest feasible baselines using exactly the same four references, 8,192 query points, input preprocessing, and object-score aggregation? For results imported from external papers, could they enumerate protocol differences that may influence each comparison?

Answer1: Under a matched evaluation protocol, MARS stays first on both metrics across both benchmarks, 4.26–7.28 pp ahead of the strongest baseline for each metric. We reran MARS and the four registration-based baselines on the same queries, used the same four normal scans where applicable, subtracted the full-cloud centroid where compatible (following Reg3D-AD and PASDF’s Real3D-AD protocol), and scored the same 8,192 points sampled by MARS using its top-16 harmonic mean for pooling. Where a method needs its own pipeline, we kept it: Reg2Inv estimates pose and features jointly, and PASDF uses its official template and pretrained SDF checkpoint; we read their scores at the same points.

Table R3. Results reported in Table 1 and rerun under the matched protocol (P/O-AUROC, %).

MethodReal3D-AD, Table 1Real3D-AD, matchedAnomaly-ShapeNet, Table 1Anomaly-ShapeNet, matched
PointCore [G5]73.1 / 82.975.76 / 80.62— / 67.766.23 / 67.06
Reg3D-AD [G1]70.0 / 70.465.92 / 59.5166.8 / 57.255.18 / 54.37
PASDF [G4]74.5 / 80.275.08 / 81.8989.7 / 90.089.54 / 89.75
Reg2Inv [G3]87.8 / 78.090.91 / 80.0788.2 / 86.193.11 / 87.67
MARS (Ours)97.0 / 91.295.17 / 89.1798.2 / 95.798.25 / 95.05
  • Matching the protocol changes individual numbers but not the conclusion. It reorders some baselines (PASDF now exceeds PointCore in Real3D-AD O-AUROC) and also lowers MARS on Real3D-AD. We did not retune any method or hyperparameter for this protocol.
  • The matched protocol does not favor MARS. On Real3D-AD, subtracting the full-cloud centroid before alignment removes the sensor-to-centroid direction that MARS uses to orient depth maps for partial scans, which can affect pose initialization; MARS still stays first.

Answer2: The imported results in Table 1 keep each paper's own protocol: they differ in which points enter P-AUROC, how object scores are pooled, and which preprocessing and pretrained models are used. Both benchmarks give every method the same four normal training scans per category [G1, G2], although PASDF uses the provided class-specific normal template for alignment and a trained SDF for scoring (Table C.3). Table C.3 lists these settings for MARS and the four rerun baselines.

We thank the reviewer for this request. We report the matched reruns in Table C.4 and document the method-specific evaluation settings in Table C.3 (App. C.3 of the revised PDF).

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global Response [2/2]

Other revisions of the manuscript

We have highlighted newly added text and the caption titles of new or revised tables and figures in blue.

  1. Page 2, Lines 89–106 (Sec. 1), and Page 10, Lines 529–539 (Sec. 6). We restate the scoring stage, the contributions, and the conclusion around spatially aggregated residuals calibrated against normal-reference variation, to address concerns from reviewers. [ACHj, 6tbY, v4fE]
  2. Page 7, Lines 350–353 (Sec. 4.1). We explain how the shared configuration was selected on Real3D-AD. [ACHj]
  3. Page 7, Lines 374–377 and Page 25, Lines 1341–1345, and Page 26, Lines 1368–1373 (Table E.1(c)). We report the mean, SD, and range across all six two-reference pairs. [ACHj, v4fE]
  4. Page 9, Lines 447–455 and Lines 471–473 (Table 2). We report setup time, online and amortized latency, and peak CPU/GPU memory. [ACHj]
  5. Page 9, Lines 457–468 and Lines 475–480 (Table 3). We move the Real3D-AD region-wise residual analysis into the main paper. [6tbY]
  6. Pages 9–10, Lines 484–493 and Lines 505–506, and Page 19, Lines 972–981 and Lines 999–1004 (Tables 4 and D.1(a)). We add a no-viewpoint-search ablation on both benchmarks. [ACHj]
  7. Page 10, Lines 486–493 and Lines 508–511, and Page 19, Lines 983–989 and Lines 1006–1009 (Tables 5 and D.1(b)). We re-evaluate the scoring ablations with median smoothing reserved for the final step. [ACHj]
  8. Page 10, Lines 495–502 and Lines 513–517, and Page 23, Lines 1198–1208 and Lines 1224–1236 (Tables 6 and D.7(b–c)). We isolate normal calibration with fusion fixed, distinguish raw-residual from probability-based scoring, and add Mahalanobis aggregation. [ACHj, 6tbY, v4fE]
  9. Page 18, Lines 918–966 (Tables C.3–C.4). We document the baseline protocols and report matched reruns alongside the original results. [ACHj]
  10. Pages 19–20, Lines 1011–1064 (Table D.2). We add a cumulative scoring ablation with 95% paired bootstrap confidence intervals. [ACHj]
  11. Pages 20–21, Lines 1066–1070 and Lines 1080–1088 (Table D.3). We report ICP iteration budgets and observed mean counts with early stopping. [ACHj]
  12. Pages 20–21, Lines 1072–1079 and Lines 1090–1110 (Table D.4). We compare single-scale and multi-scale scoring and quantify the effect of standard-deviation channels. [6tbY]
  13. Pages 23–24, Lines 1238–1257 (Table D.8). We report category-wise calibration gains, ties, and declines. [6tbY, v4fE]
  14. Pages 24–25, Lines 1278–1284 and 1326–1333 (App. D.10, Table D.10). We compare viewpoint-preserving candidate selection with global top-1, top-2, and top-4 selection. [ACHj]
  15. Pages 26–28, Lines 1390–1402 and 1427–1430, and Lines 1449–1452 and 1490–1493 (Figs. E.1–E.2). We show two failure cases involving pose mismatch and normal shape variation. [6tbY]
  16. Pages 28–29, Lines 1458–1469 and Lines 1495–1520 (Tables E.3–E.4). We report AUROC by defect type and analyze defect-point counts and two-layer patterns. [v4fE]

We hope these revisions address all concerns and further clarify the significance of this work.

Global references

  • [G1] Real3D-AD: A Dataset of Point Cloud Anomaly Detection, Liu et al. 2023
  • [G2] Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and a Self-Supervised Learning Network, Li et al. 2024
  • [G3] Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection, Yu et al. 2025
  • [G4] Bridging 3D Anomaly Localization and Repair via High-Quality Continuous Geometric Representation, Zheng et al. 2025
  • [G5] PointCore: An Efficient Framework for Unsupervised Point Cloud Anomaly Detection Using Joint Local-Global Features, Zhao et al. 2026
  • [G6] Back to the Feature: Classical 3D Features are (Almost) All You Need for 3D Anomaly Detection, Horwitz & Hoshen 2023
  • [G7] Statistical Methods for Research Workers, Fisher 1925
  • [G8] PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection, Ye et al. 2025
  • [G9] Probing 3D Anomalies via Multi-View Registration and Dual-Residual Analysis (MVR-DR), Yang et al. 2026
AuthorsOpenReview ↗

Global Response [1/2]

We sincerely thank all reviewers (ACHj, 6tbY, v4fE) for their careful reviews and valuable feedback, which have substantially improved our work. Below, we summarize the strengths the reviewers highlighted, restate our main contributions, and highlight the key responses and new results. The other revisions of the manuscript and the references cited in our responses follow in the next comment.

Strengths of our work, as highlighted by reviews:

  1. [ ACHj, 6tbY, v4fE ] Training-free detection from only a few normal references, without pretrained networks, anomalous examples or synthesized defects.
  2. [ ACHj, 6tbY, v4fE ] Strong performance on both real partial scans and complete synthetic shapes, outperforming the reported baselines on average.
  3. [ 6tbY ] Insightful multi-scale residual statistics that consistently improve localization.
  4. [ ACHj, v4fE ] Thorough ablations, robustness tests and runtime comparisons that make the source of performance credible.
  5. [ v4fE ] A simple and effective design that detects anomalies directly from geometric differences.

Main contributions of our work

  • A finding: once a query is aligned with the normal scans, the distance from each query point to the nearest normal point (the residual) is a strong anomaly signal on its own. Built only from classical tools and without training, MARS achieves the highest four-shot P- and O-AUROC on both benchmarks (Table 1).
  • Viewpoint-preserving alignment keeps a partial scan from being locked into a wrong pose. MARS keeps the best pose candidate from each viewing direction until the final refinement.
  • Residuals are summarized over several neighborhood sizes and judged against how they vary among the normal scans. Each of the seven multi-scale statistics is scored by how rare it is among the normal scans, and Fisher-style fusion combines them.

Highlighted responses and new results

  • [ACHj] Under one shared evaluation protocol, MARS stays first on every metric, 4.26–7.28 pp ahead of the strongest of four rerun baselines.
  • [ACHj] The configuration selected on Real3D-AD transfers to Anomaly-ShapeNet without adjustment. Across both transfer directions, the largest loss was 0.42 pp in O-AUROC. In the hyperparameter sensitivity analysis, varying one setting at a time changed P- and O-AUROC by at most 0.23 and 1.14 pp, respectively.
  • [ACHj, v4fE] MARS is robust to the choice of normal references. Across all six two-reference pairs, the standard deviation of P- and O-AUROC is at most 0.70 pp on either benchmark.
  • [6tbY, v4fE] Comparing residuals with the normal scans adds to the multi-scale statistics: it raises P/O-AUROC by 1.10/7.54 pp on Real3D-AD and 0.53/1.10 pp on Anomaly-ShapeNet over ranking them within the query, most for object-level detection on partial scans.
  • [ACHj] MARS stays fast when its one-time preparation is counted. 1.36 s per query, less than the online inference alone of the registration-based baselines (4.12–17.46 s).
  • [ACHj] A cumulative scoring ablation quantifies the gains from multi-scale statistics, normal LOO calibration, and Fisher-style fusion, with 95% paired confidence intervals for each step (Table D.2).
  • [ACHj] Keeping one candidate per viewpoint pays off on partial scans: compared with global top-4 selection, it raises O-AUROC by 3.24 pp on Real3D-AD, with no shortlist size to tune (App. D.10).
  • [6tbY, v4fE] We illustrate two possible contributors to false alarms and report results for selected defect types. The cases show how pose mismatch and normal shape variation can contribute to false alarms. Defects with fewer labeled points and more frequent two-layer patterns generally yield lower AUROC for both MARS and the baselines.

We hope these revisions address all concerns and further clarify the significance of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.