acceptodds
Withdrawn from ICLR 2027

GMetric-3D: Geometric Metric Encoding Based on Topological Consistency for 3D Industrial Defect Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The 3D industrial surface defect detection is a challenging task due to the small scale, irregular shape, and lack of stable semantic patterns of the defects. However, the existing 3D defect detection methods mainly rely on semantic and appearance features and fail to effectively capture these subtle geometric variations. To address this issue, we propose the geometric metric encoding based on topological consistency (GMetric-3D) that explicitly models the local geometric structure in defects. Specifically, we first introduce the geometric metric encoding to capture local structural patterns by combining metrics such as relative positions, normal vector variations, and curvature information within the point neighborhood. To further enhance the distinguishability of defect features, we present the topological consistency learning, which retains topological structure in normal regions and expands feature differences in defect regions through topological smoothing, separation, and boundary perception constraints. Extensive experiments on the high-precision 3D point cloud dataset CPS3D-Det and the large-scale scene dataset KITTI demonstrate that our GMetric-3D achieves state-of-the-art performance compared with existing 3D detection methods. The experimental results confirm that modeling geometric structure information is crucial for reliable and accurate 3D industrial defect detection. The source code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geometric metric encoding, topological consistency learning, geometric variations, industrial defect detection

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_kRjsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes GMetric-3D, a 3D industrial defect detection framework that aims to improve the detection of subtle defects by explicitly modeling local geometric structures. The method introduces a geometric metric encoding that combines relative point positions, normal-vector variations, and curvature information within local neighborhoods. It further employs topological consistency learning to preserve feature consistency in normal regions while increasing feature differences around defect regions through topological smoothing, separation, and boundary-perception constraints.

The authors claim that GMetric-3D achieves state-of-the-art performance compared with existing 3D detection methods. The method is evaluated on the CPS3D-Det high-precision 3D point-cloud dataset and the KITTI large-scale scene dataset, with quantitative experimental results reported to demonstrate the effectiveness of the proposed geometric representation and topological consistency learning.

Critical strengths
  1. Relevant and practically important problem. The paper addresses 3D shape anomaly detection, which is an important problem with potential applications across various 3D vision and inspection domains. The topic is timely and practically relevant, providing a strong motivation for the proposed work.
  2. Effective use of complementary geometric features. The proposed method jointly leverages curvature features together with surface normals and neighboring point coordinates to characterize local geometric structures. Combining these complementary geometric cues is a promising approach for identifying local shape anomalies, as it can capture both surface properties and local spatial context more effectively than relying on a single feature type.
Critical weaknesses
  1. Insufficient experimental validation to support the paper’s claims. The current evaluation is too limited to convincingly demonstrate the effectiveness and generalizability of the proposed method. The experiments appear to rely primarily on the Real3D-AD dataset, while other relevant datasets, such as Shape-Anomaly, are not evaluated. In addition, the qualitative evaluation is very limited: the paper does not provide sufficient whole-shape anomaly detection visualizations, and Fig. 1 mainly shows local feature visualization rather than complete detection results. More comprehensive quantitative and qualitative comparisons, including results on diverse datasets and representative anomaly cases, are necessary to substantiate the claimed improvements.

  2. The technical novelty and contribution are not sufficiently clear. The proposed framework appears to combine neighboring point coordinates, surface normals, and curvature features, but the paper does not clearly explain what constitutes the novel methodological contribution beyond combining existing geometric cues. Similar geometric features and local point/normal information have already been widely explored in 3D shape anomaly detection methods, including PO3AD and hierarchical point-patch-based approaches. In particular, the current presentation does not clearly distinguish which components are learned versus non-learned, nor does it sufficiently explain how curvature is explicitly incorporated and contributes to anomaly detection. A detailed ablation study is also missing to isolate the contribution of neighboring point positions, normal directions, and curvature features, ideally with both quantitative results and corresponding qualitative anomaly visualizations. Without this evidence, it is difficult to determine whether the reported performance gains arise from a genuinely novel methodological contribution or primarily from combining existing geometric features.

Missing important related references:

[1] X. Kang, Z. Li, T. Lan, D. Gong, K. Khoshelham, and L. Nan, “Hierarchical point-patch fusion with adaptive patch codebook for 3D shape anomaly detection,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2026, pp. 24258–24267. [2] Ye J, Zhao W, Yang X, et al. Po3ad: Predicting point offsets toward better 3d point cloud anomaly detection[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 1353-1362.

Questions for authors
  1. What is the precise methodological novelty of the proposed approach compared with existing 3D shape anomaly detection methods such as PO3AD and hierarchical point-patch-based approaches? In particular, please clarify which components are newly proposed and which are adopted or combined from existing methods, and explain how the proposed use of curvature provides information beyond the point-coordinate and surface-normal features already used.

  2. Can the authors provide a more complete experimental justification for the claimed performance? In particular, why is the evaluation limited primarily to Real3D-AD while other relevant datasets, such as Shape-Anomaly, are not considered? Please clarify whether technical or data-related reasons explain this limitation, and whether the reported conclusions hold across different datasets and anomaly types.

  3. How much does each geometric feature contribute to the final performance? Specifically, what are the individual and combined contributions of neighboring point coordinates, surface normals, and curvature? A detailed ablation study with both quantitative metrics and qualitative whole-shape anomaly visualizations (complete visual results for qualitative evaluation are missing) would help determine whether the reported gains are attributable to the proposed method rather than simply to combining multiple existing geometric features.

Reviewer_CBb6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GMetric-3D, a supervised 3D detector for localizing small industrial surface defects with 3D bounding boxes. The method builds on the sparse convolutional backbone and detection pipeline of AIFAD. Its geometric metric encoding module combines relative voxel offsets, differences between learned vectors described as normals, and a PCA-based local curvature statistic. The resulting neighborhood features are fused with the backbone representation. The paper then adds three topological consistency losses intended to smooth normal regions, separate structurally inconsistent regions, and preserve defect boundaries.

The method is evaluated on the CPS3D-Det and KITTI validation sets. On CPS3D-Det, it improves over AIFAD on three of the four reported AP columns, although the gains are modest and it is worse on R40 BEV AP. The paper also reports several configuration studies, runtime measurements, training curves, and feature visualizations.

Critical strengths
  1. The paper studies a concrete and relatively underexplored task: supervised 3D bounding-box detection of small industrial surface defects. It also clearly distinguishes this setting from normal-only anomaly detection. CPS3D-Det is an appropriate benchmark for the task, and reporting both 3D and BEV AP is useful.

  2. The geometric module is straightforward to integrate into a sparse detector and remains fast in the reported setup. The comparison covers a broad set of standard point, voxel, and point-voxel detectors, and the full model is competitive on CPS3D-Det. These are useful engineering results, although they do not overcome the methodological problems below.

Critical weaknesses
  1. The mathematical formulation of topological consistency is insufficiently justified and contains several ambiguities. Equation 9 applies smoothing to all neighboring point pairs without explicitly restricting the loss to normal regions. The inconsistent-pair set SS in Equation 10 is not clearly defined, making the separation objective difficult to reproduce. More importantly, Equation 11 identifies boundary points based on learned feature distances, while Equation 12 minimizes a positively weighted sum of these distances. This appears to encourage smaller feature differences at the selected boundaries, which is difficult to reconcile with the stated goal of preserving structural discontinuities. The geometric encoder also raises a related concern: nin_i is described as a surface normal, but it is obtained through a learned projection of FiF_i, without a clearly specified normal supervision objective, normalization constraint, or orientation convention. These issues affect the interpretation and validity of the proposed core components.

  2. The experimental evaluation does not adequately isolate the proposed contributions or support several of the paper's claims. Table 3 does not include the complete model, and its last three configurations vary a geometric input together with a TCL loss, making their individual contributions difficult to determine. Table 4 is presented as a backbone ablation but compares the complete GMetric-3D system against different detection architectures, rather than changing only the backbone within a controlled framework. No repeated-run statistics or uncertainty estimates are provided. On CPS3D-Det, the improvements over AIFAD are only 1.19 and 0.67 percentage points in R11 3D AP and BEV AP, respectively, while R40 BEV AP decreases from 38.41 to 37.20. Furthermore, the abstract's claim of state-of-the-art performance on KITTI is not consistently supported by Table 2, where several existing methods achieve better results across multiple metrics.

Questions for authors
  1. Clarification: What are the exact implemented definitions of the normal-region pairs used by Lsmooth\mathcal{L}{\mathrm{smooth}}, the inconsistent-pair set SS used by Lsep\mathcal{L}{\mathrm{sep}}, and the boundary points used by Lboundary\mathcal{L}_{\mathrm{boundary}}? As written, the smoothing loss has no explicit normal-region mask, SS is insufficiently defined, and minimizing Equation 12 appears to penalize large feature differences at selected boundaries. Please clarify the implementation and explain how the boundary objective preserves rather than suppresses structural discontinuities.

  2. Clarification: Is nin_i in Equation 4 explicitly trained to represent a geometric surface normal? If so, what supervision, normalization, and orientation conventions are used? Otherwise, what justifies interpreting differences between learned feature projections as normal-vector variations?

  3. Can the authors provide a controlled ablation including the baseline, GME only, TCL only, the full model, and one-at-a-time removal of each geometric metric and loss component, preferably across multiple random seeds? Please also clarify why the full model is absent from Table 3 and reconcile the broad state-of-the-art claim on KITTI with the mixed results reported in Table 2.

Reviewer_vEAMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper works on 3D industrial defect detection, where defects are located with 3D boxes in high-resolution point clouds. The authors add two parts to a voxel-based detector. The first is a geometric metric encoding, which takes relative positions, normal differences and curvature in a grid neighborhood and fuses them with the backbone features. The second is a topological consistency learning module, which uses three extra losses for feature smoothing, separation and boundaries. Experiments are done on CPS3D-Det and KITTI, with ablations on the modules, backbone, loss weights, neighborhood size, fusion and detection head.

Critical strengths
  1. The task is practical, and treating defects as local geometric changes instead of semantic objects makes sense.
  2. The method is simple and runs fast.
  3. Many detectors are compared, with params, FLOPs and latency.
Critical weaknesses
  1. The novelty is limited. The curvature is the standard surface variation, the encoder is similar to EdgeConv/PointNet++, and the separation loss is a usual contrastive margin loss.
  2. The abstract claims SOTA on KITTI, but GMetric-3D is not the best in any column of Table 2.
  3. The efficiency numbers look strange. The model has 16× more FLOPs than AIFAD but is reported faster, and SECOND’s latency is 113 ms in Table 1 vs. 39 ms in Table 4
  4. The gain over AIFAD is small (about +1.2 3D AP, and BEV AP R40 is lower),and Fig. 5 shows large jumps between epochs.
  5. The ablations don’t match the main result. Table 3 has no full-model row, and its best row is 29.39 while the main result is 40.90. The baseline is 22.74, but AIFAD, whose backbone is used, already gets 39.47 in Table 1. Tables 4 and 5 also don’t reach the main numbers.
Questions for authors
  1. What is the full-model result in Table 3, and why is the baseline far below AIFAD with the same backbone? Which λ values are used for the main result?
  2. How are the normals supervised, and how is the pair set S in Eq. 10 built?
  3. Why is GMetric-3D faster than AIFAD with much higher FLOPs?

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

Related papers

Discussion (0)

Sign in to comment.