acceptodds
Withdrawn from ICLR 2027

Batched Point Cloud Anomaly Detection via Structural Position-aware Comparison

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Point Cloud Anomaly Detection (PCAD) is critical for industrial inspection. We propose Structural position-aware point cloud Anomaly Detection (SpotAD), the novel batched PCAD framework that performs anomaly detection directly on point cloud. Our SpotAD enables position-aware comparison through deterministic PCA-based alignment. To ensure both efficiency and structural consistency, we introduce object-aware prototype selection to restrict comparisons to structurally similar objects, position-aware prototype filtering for geometrically constrained comparison, and local prototype score smoothing to propagate prototype-level anomaly scores to local neighborhoods, enforcing spatial continuity. This strategy preserving geometric correspondence. Extensive evaluations on four benchmark datasets demonstrate state-of-the-art performance under the batched PCAD setting, achieving 87.0% object-level AUROC and 86.3% point-level AUROC, while outperforming existing zero-shot and full-shot PCAD methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point cloud anomaly detection, training-free batched, Structural position-aware

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vaM7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SpotAD, a training-free batched point cloud anomaly detection method. The method identifies anomalies by exploiting shared normal geometric structures across test samples. Specifically, it first employs PCA for point cloud pose alignment and constructs local geometric prototypes using FPFH descriptors and FPS sampling. Subsequently, Object-aware Prototype Selection (OPS) retrieves structurally similar reference objects, while Position-aware Prototype Filtering (PPF) imposes spatial constraints on local geometric feature matching. Local Prototype Score Smoothing (LPSS) is further introduced to improve the spatial consistency of anomaly localization. Experiments on four benchmarks, including Anomaly-ShapeNet, Real3D-AD, MulSen-AD, and MiniShift, achieve an average object-level AUROC of 87.0% and point-level AUROC of 86.3%.

Critical strengths
  1. This paper introduces a batched anomaly detection paradigm for native 3D point clouds, enabling training-free anomaly detection through cross-sample comparisons of local geometric structures. By avoiding category-specific training and multi-view rendering, the method offers certain practical advantages.
  2. The proposed method achieves competitive detection and localization performance across four 3D anomaly detection benchmarks. The authors also conduct component ablations, batch composition analysis, and comparisons between transductive and inductive settings to validate the effectiveness of their approach.
Critical weaknesses
  1. The methodological novelty of this work is relatively limited. PCA-based pose alignment, FPFH feature extraction, nearest-neighbor prototype matching, and neighborhood-based score smoothing are all established techniques. OPS and PPF essentially perform nearest-neighbor candidate selection based on object-level feature similarity and spatial positional constraints, respectively. Therefore, the overall framework mainly combines existing geometric processing and retrieval strategies. Moreover, prior studies such as MuSc have already explored batched anomaly detection through mutual comparisons among test samples. The paper does not sufficiently clarify its substantive methodological contributions beyond existing batched anomaly detection approaches.
  2. The current experiments do not sufficiently establish the necessity of the key components. Table 3 shows that removing OPS results in average O-AUROC and P-AUROC scores of 86.0% and 86.2%, respectively, compared with 87.0% and 86.3% for the full model, indicating relatively limited improvements. Furthermore, the paper lacks comparisons with simple cross-sample matching baselines, such as PCA+FPFH+KNN, as well as alternative batched detection strategies under the same test-sample reference setting. Consequently, it remains unclear whether the performance gains primarily arise from the batched reference mechanism or from the proposed position-aware prototype selection strategies.
  3. The reliability of geometric alignment and spatial correspondence is insufficiently validated. SpotAD relies on PCA to establish a common coordinate system across point clouds. However, PCA may fail to provide reliable alignment for rotationally symmetric objects, objects with unstable principal axes, or point clouds containing substantial local deformations. Moreover, PPF performs matching based only on the normalized direction of prototype centers relative to the object center, without considering radial distance information, potentially leading to incorrect correspondences between distinct local structures. Given that Table 3 shows a substantial decrease in average O-AUROC from 87.0% to 59.2% when PCA is removed, further analysis is needed to assess the robustness of this alignment-dependent mechanism under challenging geometric conditions.
  4. The experimental evaluation provides insufficient evidence regarding the applicability and computational efficiency of the method. SpotAD constructs its reference bank from test samples, which represents a different inference setting from zero-shot methods that process individual test samples independently. Therefore, direct performance comparisons may not be entirely fair. Furthermore, although the paper emphasizes efficiency and scalability, it lacks comprehensive evaluations of end-to-end inference time, GPU memory consumption, and computational costs under different reference bank sizes. These omissions weaken the empirical support for the claimed efficiency and scalability.
Questions for authors
  1. SpotAD primarily combines established techniques such as PCA alignment, FPFH descriptors, and nearest-neighbor matching. What are the key methodological innovations compared with conventional geometric matching approaches?
  2. Table 3 shows that removing OPS decreases the average P-AUROC by only 0.1 percentage points. How do the authors justify the necessity of OPS given its limited performance contribution?
  3. PCA alignment may be unstable for symmetric objects, varying rotations, and local geometric defects. How do the authors ensure that the spatial correspondences required by PPF remain reliable under these conditions?
  4. When multiple anomalous samples contain similar defects at the same spatial locations, could cross-sample prototype matching incorrectly identify these recurring anomalies as normal structures?
Reviewer_NKZROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes SpotAD, a training-free framework for batched point cloud anomaly detection. The method uses PCA-based alignment to establish geometric correspondence, followed by Object-aware Prototype Selection (OPS), Position-aware Prototype Filtering (PPF), and Local Prototype Score Smoothing (LPSS) for anomaly scoring. Experiments on four 3D anomaly detection benchmarks report average object-level and point-level AUROC scores of 87.0% and 86.3%, respectively. Additional experiments examine the effects of individual components, evaluation protocols, and batch composition.

Comment

The paper presents a simple and effective approach to training-free point cloud anomaly detection, with promising results across several benchmarks. However, the lack of a directly comparable batched baseline makes it difficult to assess the method's contribution, while the robustness of its geometric alignment and matching assumptions remains insufficiently validated. I lean toward rejection but would reconsider if these concerns are addressed.

Critical strengths

1. Simple and practical framework. SpotAD operates directly on point clouds without model training or auxiliary annotations. The position-aware comparison strategy is intuitive and avoids the information loss associated with 2D projection.

2. Strong experimental results. The method achieves competitive performance across four benchmarks, outperforming the reported baselines on average. The ablation studies and batch-composition experiments provide useful evidence of its effectiveness.

Critical weaknesses

1. Insufficient comparison under the same batched setting.

The main comparisons are against full-shot and zero-shot methods, which have different data access and inference protocols. Since SpotAD builds its reference library from the test batch, these comparisons do not fully establish the advantage of the proposed method. A direct comparison with a simple 3D adaptation of MuSc, using the same point features and test-batch reference pool, would be more convincing.

2. Limited validation of alignment and batch robustness.

The method relies heavily on PCA alignment and position-aware matching, but their robustness to symmetric objects, ambiguous principal axes, and large pose variations is not systematically evaluated. Moreover, although Table 2 studies different anomaly ratios, the effects of small batches and severe class imbalance remain unclear. These factors may limit practical deployment.

Questions for authors

Q1. Can the authors compare SpotAD with a MuSc-style batched baseline operating directly on point clouds under the same evaluation protocol?

Q2. How robust is the PCA-based alignment to symmetric shapes, ambiguous principal directions, and large rotations? Can the authors provide a quantitative analysis?

Q3. How does performance change with batch size and class imbalance? Since efficiency is a key motivation, can the authors also report inference time and memory usage against the baselines?

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

This paper studies batched point cloud anomaly detection (Batched Point Cloud Anomaly Detection) that requires no training data. The authors propose SpotAD, which directly computes object-level and point-level anomaly scores through structural comparison among test samples, thereby avoiding the reliance of traditional unsupervised PCAD on normal training samples and the possible loss of geometric information caused by 2D projection-based methods. The method consists of four main components: PCA-based canonical pose alignment, Object-aware Prototype Selection (OPS), Position-aware Prototype Filtering (PPF), and Local Prototype Score Smoothing (LPSS). The authors report an average O-AUROC of 87.0% and P-AUROC of 86.3% on four datasets, demonstrating competitive performance.

Critical strengths

This paper studies a problem of important practical significance, namely performing anomaly detection directly on 3D point clouds without category-specific training and without relying on auxiliary annotations. The proposed method combines PCA-based alignment, object-aware prototype selection, and position-aware prototype filtering, providing a clear and reasonable geometric motivation for detecting local structural deviations. The proposed method achieves strong average performance on four benchmark datasets.

Critical weaknesses

. The paper's claimed contribution does not hold.

The paper claims to propose the first novel batched framework for point cloud anomaly detection using only 3D data. However, MuSc-V2 explicitly supports 3D-only point cloud inputs and performs anomaly detection through mutual scoring among unlabeled samples. Its official implementation also provides a 3D-only mode using Point-MAE. Although the authors discuss the differences between their method and MuSc-V2 in Section 2.2, they do not include MuSc-V2 in the experimental comparisons or report its performance.

  1. The experimental comparisons are insufficient. Several relevant methods are discussed in the paper and even included in the references, but are not included in Table 1 for comparison. Examples include MuSc-V2 and Reg2Inv (NeurIPS 2025). Furthermore, many recently published methods are also missing from the comparisons, such as BTP (CVPR 2026) and AARD (CVPR 2026), among others. We are not raising these papers due to any personal interests or conflicts of interest. Rather, given the contributions claimed by the authors, these methods, including but not limited to those mentioned above, should be comprehensively compared to substantiate the paper's claims.
Questions for authors

Q1. Novelty and distinction from MuSc-V2.Given that MuSc-V2 already supports 3D-only point cloud anomaly detection through mutual scoring of unlabeled samples, how do the authors justify the claim that SpotAD is the first batched point cloud anomaly detection framework using only 3D data? What are the fundamental differences between SpotAD and the 3D-only setting of MuSc-V2, and what constitutes the main novel contribution beyond this existing framework?

Q2. Completeness of experimental comparisons.Why are several closely related and already published methods, particularly MuSc-V2 (TPAMI 2026) and Reg2Inv (NeurIPS 2025), which are already discussed or cited in the paper, omitted from Table 1? Furthermore, why are recent methods such as BTP (CVPR 2026) and AARD (CVPR 2026) not included? Could the authors provide comparisons with these methods under consistent evaluation settings, especially the 3D-only version of MuSc-V2, or provide concrete justifications for their omission?

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

Related papers

Discussion (0)

Sign in to comment.