Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Towards High-Resolution 3D Anomaly Detection: A Scalable Dataset and Real-Time Framework for Subtle Industrial Defects

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims subtle industrial defects are best detected in 500k-point clouds, and its MiniShift benchmark plus Simple3D detector outperform prior methods on it and three existing benchmarks at over 20 FPS.

desk verdict A genuinely useful high-resolution 3D anomaly detection benchmark with a strong simple baseline, but the paper's “high-resolution processing” claim is undercut by the fact that the method actually runs on a few thousand sampled points and the synthetic anomaly realism is not validated. read the letter →

arxiv 2507.07435 v1 pith:DC2MTFRG submitted 2025-07-10 cs.CV

Yuqi Cheng , Yihan Sun , Hui Zhang , Weiming Shen , Yunkang Cao This is my paper
classification cs.CV
keywords 3Danomalydetectionhigh-resolutionpointcloudindustrialdefectsyntheticsynthesisunsupervisedFPFHdescriptorreal-timeinferencedifficultytiers
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Most 3D anomaly detection benchmarks and methods operate on heavily downsampled point clouds of around 1,000 points, which washes out the small surface defects that matter in industrial inspection. The paper argues that high-resolution input, roughly 500,000 points per object, is what makes subtle defects visible, and that existing methods cannot afford to process that resolution with learned backbones. To test this, the authors build MiniShift, a synthetic 500k-point benchmark whose anomalies occupy less than 1% of the surface, and Simple3D, a detector built on multi-scale handcrafted neighborhood descriptors with local spatial aggregation. They report that Simple3D surpasses prior methods on MiniShift and on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD, while sustaining real-time inference above 20 FPS. The paper's case is that high-resolution geometry plus efficient local features is the practical route to subtle defect detection.

What carries the argument

Three pieces carry the argument. AG-GAS is a parameterized geometric synthesis framework that turns anchor-point selection, Dijkstra shortest-path computation, radius-based region expansion, and normal-direction scaling into controlled protrusions and depressions, generating diverse defect shapes instead of only circular bumps or dents. Multi-Scale Neighborhood Description (MSND) concatenates FPFH handcrafted descriptors computed over several neighborhood sizes per point, capturing local geometry at complementary scales. Local Feature Spatial Aggregation (LFSA) averages those multi-scale descriptors over sampled neighborhoods, widening each feature's receptive field at low computational cost. Detection is prototype-based: normal training clouds provide a set of aggregated features, each test point is scored by its distance to the nearest prototype, and the object score is the maximum point score. The paper's argument is that these efficiently computable local features preserve the small geometric cues that downsampling and learned tokenization destroy.

What would settle it

Scan a set of real industrial parts with human-confirmed micro-defects at 500k points, compare the geometric statistics of the real defects with those of AG-GAS's four types, and measure Simple3D's point-wise AUROC on those real defects; if the real-defect AUROC falls toward chance while MiniShift AUROC stays high, the benchmark's realism premise would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that resolution is the bottleneck: a small surface imperfection is nearly invisible at 1k points, becomes discernible at 8k points, and is clearly detectable at 500k points. The paper introduces a scalable synthesis pipeline, Anchor-Guided Geometric Anomaly Synthesis (AG-GAS), which selects two anchor points, computes a geodesic path between them along the point cloud graph, expands that path into a mask region, and deforms the region along its average normal with a displacement that fades with distance from the path. This produces four defect types, Areal, Striate, Scratch, and Sphere, at three difficulty tiers controlled by length, radius, and depth parameters. The resulting MiniShift dataset contains 2,577 point clouds at 500k points per sample with anomalies under 1% of the surface area, and the paper shows that prior prototype-, reconstruction-, and regression-based methods degrade sharply on it, especially at the hard tier. Simple3D computes FPFH descriptors at multiple neighborhood scales, aggregates those descriptors over sampled neighborhoods, and scores test points by their distance to normal prototypes; the paper reports it achieves the best object-wise and point-wise AUROC on MiniShift, Real3D-AD, Anomaly-ShapeNet, and MulSen-AD while exceeding 20 FPS at 4k points.

Load-bearing premise

The load-bearing premise is that the smooth, geodesic-shaped distortions produced by AG-GAS are representative of real subtle industrial defects; if real machining, casting, or impact defects differ in shape, boundary sharpness, or depth distribution, then MiniShift's results may not transfer to factory scans.

Editorial extensions

If this is right

  • Detection accuracy on MiniShift and the three established benchmarks rises as input point resolution increases, and Simple3D only overtakes prior methods when resolution reaches roughly 4,000 points or more.
  • Real-time screening becomes feasible: at 4k points Simple3D runs above 20 FPS, moving high-resolution anomaly detection from offline analysis toward inline industrial inspection.
  • Learned point-cloud backbones are not required for state-of-the-art subtle-defect detection; multi-scale handcrafted descriptors plus local spatial aggregation are enough.
  • The three-tier difficulty protocol gives benchmarks a graded stress test, since methods that look competitive on easy anomalies can drop toward random performance on the hard tier, so future evaluations should report difficulty-stratified results.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not test whether AG-GAS distortions match real machined or cast defects; a direct geometric comparison between synthesized anomalies and measured scans of real subtle defects would settle how well MiniShift rankings transfer to factories.
  • AG-GAS could be repurposed as a training-data generator rather than only an evaluation tool, and if its anomaly distribution is diverse enough, detectors trained on it might generalize to unseen defect shapes.
  • A hybrid pipeline is a natural next step: use a learned method at low resolution to propose candidate regions, then apply Simple3D's descriptor-based scoring at full resolution inside those regions to balance speed and localization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces MiniShift, a high-resolution 3D anomaly detection benchmark built by applying a synthetic anomaly-generation pipeline, AG-GAS, to MulSen-AD source models. Each MiniShift point cloud contains 500,000 points and the synthetic anomalies occupy less than 1% of the surface, with anomalies organized into easy, medium, and hard difficulty tiers. The paper also proposes Simple3D, an unsupervised prototype-based detector that computes multi-scale FPFH descriptors (MSND), aggregates them over local neighborhoods (LFSA), and scores anomalies by distance to a memory bank of normal prototypes. Experiments report state-of-the-art results on MiniShift, Real3D-AD, Anomaly-ShapeNet, and MulSen-AD, together with a claim of real-time inference above 20 FPS.

Significance. If the claims hold, the paper makes a useful contribution: a public, large-scale, high-resolution benchmark for 3D anomaly detection, a simple and fast handcrafted-feature baseline, and extensive comparisons across several established benchmarks. The release of code, dataset, and homepage is a concrete strength, as is the breadth of evaluation. The result that a handcrafted descriptor with spatial aggregation outperforms learned methods is interesting and credible given the unsupervised setting. However, the central practical claims depend on two premises that are not currently established: that the method actually operates at the advertised high resolution, and that the synthetically generated anomalies are representative of real subtle industrial defects. Both are addressable but need explicit validation before the headline conclusions can be accepted.

major comments (4)
  1. [Abstract and Section "Analysis on Detection Resolution" (Figure 6)] The operating resolution is ambiguous and the headline claims are not directly supported. The abstract and introduction emphasize 500,000-point clouds and state that "high-resolution data" is pivotal, but Figure 6 varies the detection resolution t only from 256 to 8,192 points, and the runtime analysis reports >20 FPS at 4,096 points. The MiniShift experiments in Table 2 appear to use 8k points, not 500k points, as stated in the introduction. The paper must state explicitly what resolution is used for each experiment and report accuracy and FPS at the full 500k resolution if that is the claimed operating point. Without this, the abstract's "real-time inference exceeding 20 fps" and the conclusion that high resolution is pivotal overstate the evidence.
  2. [Section 3.1, AG-GAS, Eqs. (3)-(5) and Table 1] The representativeness of the synthetic anomalies for real industrial defects is not validated. AG-GAS generates smooth normal-direction displacements with a linear falloff along geodesic paths, with difficulty ranges set by hand-defined values of alpha, beta, and gamma. The paper provides no comparison of the geometric statistics of these anomalies (curvature, boundary sharpness, amplitude distributions) against real defect scans, no human perceptual study, and no cross-dataset transfer test. Because all MiniShift test anomalies are synthesized from the same MulSen-AD source models used to build normal prototypes, the benchmark may primarily measure how well methods detect this specific parametric distortion family. This is an addressable validation gap, but it is load-bearing for the practical inspection claims and should be filled or explicitly acknowledged.
  3. [Section 3.1, Eq. (3)] The definition of the mask region is logically inconsistent with the subsequent distance computation. Eq. (3) defines M^r as points p_j such that |p_j - p_i| < r for all p_i in Gamma, but Eq. (5) uses d_j = min_{p_m in Gamma} |p_j - p_m|, which corresponds to an existential condition (or a minimum over the path). The universal quantifier in Eq. (3) would make the mask empty for any path longer than r. This should be corrected to use min or "there exists" for the dataset construction to be reproducible.
  4. [Sections 3.2 and 3.3, Anomaly Detection] The paper does not specify how dense point-wise anomaly maps are produced. Simple3D randomly samples t points in LFSA and computes enhanced features for those points, but the point-wise anomaly score A in Eq. (9) is stated for test point clouds without explaining how scores are assigned to points outside the sampled set. If P-ROC is computed only on the sampled points, the evaluation protocol differs from the baselines, which produce dense maps by interpolation. If interpolation is used, the high-resolution localization claim needs qualification. Please clarify the exact evaluation and inference procedure.
minor comments (5)
  1. [Figure 6 caption] The caption writes 28 and 213 without superscripts; these should be 2^8 and 2^13 to match the axis descriptions.
  2. [Appendix headings and Table 12] There are several typos: "Tabel" appears repeatedly, "Anoamly-ShapeNet" is misspelled in the appendix headings, and "Scartch" appears in Figure 7.
  3. [Results text versus Table 3] The main text states that Simple3D achieves 88.3% O-ROC on MulSen-AD, but Table 3 reports 88.2%; please make the numbers consistent.
  4. [Evaluation statistics] The per-category test sets are small (30 defect instances per category-defect type and 10 normal test samples), but the paper reports no confidence intervals or significance tests. Given the large performance differences claimed, at least for the smaller O-ROC gaps this would help assess reliability.
  5. [Conclusion, hard-tier results] The hard-tier results in Table 10 show Simple3D at 61.6/56.3 on MiniShift-ALL, which is still close to chance for several categories; the sentence claiming that Simple3D "adeptly discerns even the most subtle and intricate anomalies" is stronger than the data warrant and should be tempered.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Simple3D is an unsupervised prototype-distance method with external baselines, and the MiniShift benchmark is independently synthesized without using the method's outputs.

full rationale

The paper's derivation chain is self-contained and contains no circular reduction. Simple3D's anomaly scores are computed as nearest-prototype distances in FPFH feature space (Eqs. 7-10), with the normal prototype set built only from normal training clouds; no equation defines an output in terms of the target anomaly labels or ground-truth masks. The new MiniShift benchmark is constructed by the AG-GAS pipeline (Eqs. 1-5), which synthesizes geometric distortions from held-out MulSen-AD models; the test anomalies are not fitted values, and the pipeline parameters are hand-set ranges (Table 1) that are not derived from Simple3D's outputs. Baselines come from external published works, and the two self-citations (GLFM TASE'25 and MC3D-AD IJCAI'25) appear only as comparison methods and as descriptive context, not as load-bearing justification for any claim. The abstract's 'pivotal role of high-resolution data' claim is supported by a resolution sweep (Figure 6), which is an empirical analysis rather than a definitional identity. The unvalidated realism of synthetic anomalies relative to real industrial defects is a potential external-validity limitation, not a circularity, and therefore does not affect the circularity score.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's empirical claims rest on the realism of the synthetic benchmark, the sufficiency of handcrafted local descriptors, and several hand-set hyperparameters. None of these reduce the reported metrics to a fitted constant; they are domain assumptions that a replicating group should test.

free parameters (4)
  • FPFH neighbor counts (k1, k2, k3) = 40, 80, 120
    Hand-chosen for the Multi-Scale Neighborhood Descriptor. The paper does not ablate these values, but detection performance depends on them.
  • Aggregation neighborhood size (kL) = 128 (default)
    Set in the implementation details. Figure 6 shows performance improves as kL grows from 8 to 128, so the default is the largest tested value.
  • Detection resolution t (number of sampled points) = Not stated as a single default; evaluated up to 8192
    The number of points used in Local Feature Spatial Aggregation. The real-time claim is based on roughly 4k points, but the exact default for each benchmark is not reported.
  • AG-GAS difficulty parameters alpha, beta, gamma = Ranges per Table 1
    These ratios control anomaly length, width, and depth for each difficulty tier and are hand-set to create the sub-1% footprint and subtlety of MiniShift.
assumptions (4)
  • domain assumption Synthetic anomalies generated by AG-GAS are representative of real industrial subtle defects.
    No external validation is provided, e.g., no comparison to real defect scans or human-perception study. The conclusions about high-resolution data for practical inspection rest on this premise.
  • domain assumption Multi-scale FPFH descriptors averaged in local neighborhoods can separate sub-1% geometric anomalies from normal geometry.
    Simple3D's core modeling premise. Ablations on the tested benchmarks support it empirically, but there is no theoretical or physical guarantee for arbitrary defects.
  • domain assumption Randomly sampling a few thousand points from the 500k-point cloud preserves enough anomaly points for detection.
    The reported experiments use at most 8192 sampled points, far below the dataset resolution. The paper does not analyze the worst-case sampling probability for the smallest anomalies.
  • standard math Dijkstra's algorithm on the kNN graph gives a meaningful geodesic path for defect synthesis.
    Used in AG-GAS Eq. 1-2 as a standard shortest-path computation; no further justification required.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards High-Resolution 3D Anomaly Detection: A Scalable Dataset and Real-Time Framework for Subtle Industrial Defects." pith.science (2026). https://pith.science/paper/DC2MTFRG

@misc{pith2026250707435,
  author       = {Pith},
  title        = {Pith review of: Towards High-Resolution 3D Anomaly Detection: A Scalable Dataset and Real-Time Framework for Subtle Industrial Defects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DC2MTFRG}},
  note         = {Machine review of arXiv:2507.07435}
}
read the original abstract

In industrial point cloud analysis, detecting subtle anomalies demands high-resolution spatial data, yet prevailing benchmarks emphasize low-resolution inputs. To address this disparity, we propose a scalable pipeline for generating realistic and subtle 3D anomalies. Employing this pipeline, we developed MiniShift, the inaugural high-resolution 3D anomaly detection dataset, encompassing 2,577 point clouds, each with 500,000 points and anomalies occupying less than 1\% of the total. We further introduce Simple3D, an efficient framework integrating Multi-scale Neighborhood Descriptors (MSND) and Local Feature Spatial Aggregation (LFSA) to capture intricate geometric details with minimal computational overhead, achieving real-time inference exceeding 20 fps. Extensive evaluations on MiniShift and established benchmarks demonstrate that Simple3D surpasses state-of-the-art methods in both accuracy and speed, highlighting the pivotal role of high-resolution data and effective feature aggregation in advancing practical 3D anomaly detection.

Figures

Figures reproduced from arXiv: 2507.07435 by the authors.

Figure 1
Figure 1. (a) Detecting subtle defects requires high-resolution point clouds to ensure precise localization and accurate identi [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. We commence by extracting 3D data from 12 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 2
Figure 2. The construction pipeline of MiniShift. The 3D models are utilized to generate various defects through the proposed AG-GAS framework. four anomaly types—namely Areal, Striate, Scratch, and Sphere—alongside their associated defect masks. This syn￾thesis is achieved through precise manipulation of anchor point positions and geometric deformation parameters. Ad￾ditionally, we implement a difficulty protocol, categorizi… view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: Framework of the proposed method, Simple3D. High-resolution point clouds undergo initial processing to derive multi-scale neighborhood descriptors. These descriptors are then spatially aggregated to yield enhanced point cloud features. Simple3D utilizes a prototype-bas…
Figure 4
Figure 4. Figure 4: Visualization of our point-wise anomaly detec￾tion results on MiniShift. From top to bottom: input point clouds, ground truths, anomaly maps. ifest a marked deterioration in performance, particularly in the most arduous scenarios, where their efficacy in both O￾ROC and…
Figure 5
Figure 5. Figure 5: Visualization of our point-wise anomaly detection results on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD. From top to bottom: input point clouds, ground truths, anomaly maps. Benchmarking Results on Existing Datasets To further ascertain the efficacy of Simple3D, we ext…
Figure 6
Figure 6. Figure 6: Influence of detection resolution t (First Row) and aggregated neighborhood point number kL (Second Row). The left four figures show the performance on datasets Real3D-AD, Anomaly-ShapeNet, MulSen-AD, and MiniShift, while the last one shows the Frames Per Second (FPS),…
Figure 7
Figure 7. Figure 7: Visualization of data in MiniShift between four defect types and 12 categories. The first row illustrates the high￾resolution point clouds, and the second row presents their corresponding ground truth annotations. Defect regions are magnified to enable clearer visual o…
Figure 8
Figure 8. Figure 8: Visualization of prediction results in MiniShift using the proposed method Simple3D and other methods. The first row is the original point clouds, while the second row is the ground truth. Subsequent rows depict various methods. Defect regions are magnified to enable c…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Physics-inspired Pseudo Anomaly Generation and Prototype Feature Guidance for 3D Anomaly Detection

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Physics-inspired multi-defect pseudo-anomaly generation plus momentum normal prototypes and difference-aware offset prediction yields leading 3D point-cloud anomaly detection on Anomaly-ShapeNet and Real3D-AD.

  2. Multi-View Reconstruction with Global Context for 3D Anomaly Detection

    cs.CV 2025-07 conditional novelty 5.0 of 10

    MVR projects high-resolution point clouds into multi-view depth images and reconstructs them with a pre-trained vision transformer, achieving state-of-the-art anomaly detection scores on Real3D-AD.

  3. A Comprehensive Survey for Real-World Industrial Defect Detection: Challenges, Approaches, and Prospects

    cs.CV 2025-07 conditional novelty 3.0 of 10

    A broad survey of industrial defect detection that structures the field by closed-set vs open-set and 2D vs 3D methods, with an emphasis on the rise of open-set anomaly detection.

Reference graph

Works this paper leans on

26 extracted references · 15 canonical work pages · cited by 3 Pith papers

  1. [1]

    Bergmann, P.; Jin, X.; Sattlegger, D.; and Steger, C. 2022. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 5: 202--213. Doi: 10.5220/0010865000003124

    doi:10.5220/0010865000003124 2022

  2. [2]

    Cao, Y.; Xu, X.; and Shen, W. 2024. Complementary pseudo multimodal feature for point cloud anomaly detection. Pattern Recognition, 156: 110761

    work page 2024

  3. [3]

    Cao, Y.; Zhang, J.; Frittoli, L.; Cheng, Y.; Shen, W.; and Boracchi, G. 2024. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, 55--72. Springer

    2024

  4. [4]

    Cheng, J.; Gao, C.; Zhou, J.; Wen, J.; Dai, T.; and Wang, J. 2025 a . MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection. IJCAI

    work page 2025

  5. [5]

    Cheng, Y.; Cao, Y.; Wang, D.; Shen, W.; and Li, W. 2025 b . Boosting Global-Local Feature Matching via Anomaly Synthesis for Multi-Class Point Cloud Anomaly Detection. IEEE Transactions on Automation Science and Engineering, 1--1

    work page 2025

  6. [6]

    Cheng, Y.; Cao, Y.; Xie, G.; Lu, Z.; and Shen, W. 2024 a . Towards Zero-shot Point Cloud Anomaly Detection: A Multi-View Projection Framework. arXiv preprint arXiv:2409.13162

    arXiv 2024

  7. [7]

    Cheng, Y.; Li, W.; Jiang, C.; Wang, D.; Xing, H.; and Xu, W. 2024 b . MVGR : Mean-Variance Minimization Global Registration Method for Multi-view Point Cloud in Robot Inspection. IEEE Transactions on Instrumentation and Measurement, 1--1

    work page 2024

  8. [8]

    L.; and Misa, T

    Frana, P. L.; and Misa, T. J. 2010. An interview with edsger w. dijkstra. Communications of the ACM, 53(8): 41--47

    work page 2010

Show all 26 references
  1. [9]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

    2016

  2. [10]

    Horwitz, E.; and Hoshen, Y. 2023. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2967--2976

    2023

  3. [11]

    Li, W.; and Xu, X. 2024. Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and A Self-Supervised Learning Network. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2024

  4. [12]

    Li, W.; Zheng, B.; Xu, X.; Gan, J.; Lu, F.; Li, X.; Ni, N.; Tian, Z.; Huang, X.; Gao, S.; et al. 2024. Multi-Sensor Object Anomaly Detection: Unifying Appearance, Geometry, and Internal Properties. arXiv preprint arXiv:2412.14592

    2024 arXiv

  5. [13]

    Liang, H.; Xie, G.; Hou, C.; Wang, B.; Gao, C.; and Wang, J. 2025. Look Inside for More: Internal Spatial Modality Perception for 3D Anomaly Detection. AAAI

    2025

  6. [14]

    Liu, J.; Xie, G.; Li, X.; Wang, J.; Liu, Y.; Wang, C.; Zheng, F.; et al. 2024. Real3D-AD: A Dataset of Point Cloud Anomaly Detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, volume 36

    2024

  7. [15]

    Pang, Y.; Tay, E. H. F.; Yuan, L.; and Chen, Z. 2023. Masked Autoencoders for 3D Point Cloud Self-supervised Learning. World Scientific Annual Review of Artificial Intelligence, 1: 2440001

    2023

  8. [16]

    Roth, K.; Pemula, L.; Zepeda, J.; Sch \"o lkopf, B.; Brox, T.; and Gehler, P. 2022. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14318--14328. Doi: 10.1109/CVPR52688.2022.01392

    2022

  9. [17]

    B.; Blodow, N.; and Beetz, M

    Rusu, R. B.; Blodow, N.; and Beetz, M. 2009. Fast point feature histograms (FPFH) for 3D registration. In 2009 IEEE international conference on robotics and automation, 3212--3217. IEEE

    2009

  10. [18]

    Wang, Y.; Peng, J.; Zhang, J.; Yi, R.; Wang, Y.; and Wang, C. 2023. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8032--8041

    2023

  11. [19]

    Ye, J.; Zhao, W.; Yang, X.; Cheng, G.; and Huang, K. 2025. Po3ad: Predicting point offsets toward better 3d point cloud anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 1353--1362

    2025

  12. [20]

    Yu, X.; Tang, L.; Rao, Y.; Huang, T.; Zhou, J.; and Lu, J. 2022. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 19313--19322

    2022

  13. [21]

    H.; and Koltun, V

    Zhao, H.; Jiang, L.; Jia, J.; Torr, P. H.; and Koltun, V. 2021. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 16259--16268

    2021

  14. [22]

    Zhou, Q.-Y.; Park, J.; and Koltun, V. 2018. Open3D: A modern library for 3D data processing. arXiv preprint arXiv:1801.09847

    2018 arXiv

  15. [23]

    Zhou, Z.; Wang, L.; Fang, N.; Wang, Z.; Qiu, L.; and Zhang, S. 2024. R3D-AD: Reconstruction via Diffusion for 3D Anomaly Detection. In European Conference on Computer Vision (ECCV)

    2024

  16. [24]

    Zhu, H.; Xie, G.; Hou, C.; Dai, T.; Gao, C.; Wang, J.; and Shen, L. 2024. Towards High-resolution 3D Anomaly Detection via Group-Level Feature Contrastive Learning. ACM Multimedia (ACM MM)

    2024

  17. [25]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  18. [26]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.