acceptodds
Under review as a conference paper at ICLR 2027

Learning Canonical Representations for Rotation-Invariant and Unified 3D Anomaly Detection

Abstract

3D anomaly detection is critical in industrial manufacturing because defects are defined by surface geometry that 2D inspection often misses. In practice, a single inspection pipeline covers many object categories, and the objects arrive in arbitrary poses. Recent unified methods train a single model for many categories, but they assume a fixed input pose and build on backbones pre-trained on external shape data. Methods that address rotation align each test sample to a reference, which is costly and does not carry over to the unified setting. Our key idea is to canonicalize before scoring, so that local geometry no longer carries the object pose. The canonicalizer must therefore keep the canonical frame unchanged when a defect is present and consistent across samples of the same category. We present a framework that learns Canonical representations for Rotation-invariant and Unified 3D anomaly detection (CRU3D-AD), which meets both conditions with an SO(3)-equivariant canonicalizer trained under a pairing loss and a prototype loss. The canonicalizer is then frozen while a feature extractor, a geometry-aware local attention module, and a point-wise scorer are trained on synthetically generated defects, using no external pre-training and no memory bank. Experiments on Real3D-AD and Anomaly-ShapeNet show the best point-level localization among unified methods on both datasets, with competitive object-level detection. The code will be made available upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Anomaly Detection, Canonical Representation, Unified Anomaly Detection

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_B6fQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CRU3D-AD, a unified 3D anomaly detector designed for multiple object categories and arbitrary input rotations. The method first predicts a canonical frame with an SO(3)-equivariant network. A pairing loss encourages a normal point cloud and its synthetically defective copy to use the same frame, while a category prototype loss encourages normal instances from the same category to align consistently. The canonicalizer is then frozen, and a DGCNN feature extractor, geometry-aware local attention module, and point-wise scorer are trained using synthetic bulges, dents, and scratches. A multi-prototype regularizer is used during detector training, and the final model requires neither external pretraining nor a memory bank.

On Real3D-AD and Anomaly-ShapeNet, CRU3D-AD obtains the best point-level AUROC among the unified methods included in the paper, with lower object-level AUROC than SeDiR. The evaluation also includes component and loss ablations, five-seed detector results, rotation tests, qualitative examples, and an efficiency comparison.

Critical strengths
  1. The paper addresses a useful gap between two lines of work that are usually considered separately: unified multi-category 3D anomaly detection and robustness to arbitrary object rotation. The pairing and category prototype losses are well motivated for anomaly detection, where a defect should not change the estimated frame and normal instances from one category should remain comparable after alignment. The resulting test-time pipeline uses one model without category labels, external shape pretraining, or memory-bank retrieval.

  2. The point-level localization results are strong, and the paper checks them from several useful angles. In the single-pass setting, CRU3D-AD improves P-AUROC over SeDiR from 80.6 to 85.5 on Real3D-AD and from 81.0 to 92.1 on Anomaly-ShapeNet. The paper also provides per-category results, five-seed variation, component and loss ablations, controls for structured anomaly synthesis and local attention, rotation tests, qualitative examples, and end-to-end efficiency measurements. This makes the main localization result credible.

Critical weaknesses
  1. The evaluation of arbitrary-pose robustness is narrower than the central claim requires. In each of the five runs, the same random SO(3) rotation is applied to the entire test set, rather than sampling an independent pose for each object. More importantly, the only baseline in this experiment is the authors' reproduction of MC3D-AD, which was not designed to be rotation invariant. Reg2Inv and PASDF are the closest rotation-aware alternatives, but neither is included in the shifted-pose comparison. Table B shows that CRU3D-AD itself is stable, which is encouraging, but it does not establish its advantage over the strongest relevant alternatives under arbitrary poses.

  2. The experiments do not fully isolate the value of learning the canonical frame. A random, untrained equivariant canonicalizer already obtains 85.4 P-AUROC on Real3D-AD, compared with 85.5 for the full model, although the learned model gives clearer gains on object-level AUROC and on Anomaly-ShapeNet. The paper also acknowledges that its losses do not resolve symmetric shapes. Without a direct measure of frame consistency across instances, frame stability between normal and defective pairs, or performance broken down by symmetry, it is difficult to tell when the learned canonicalizer is essential and when architectural equivariance alone is doing most of the work.

Questions for authors
  1. Decision-relevant evaluation request: Can you evaluate the case in which each test object receives an independently sampled SO(3) rotation and include Reg2Inv and PASDF, or another strong rotation-aware method, under the same protocol? Please report absolute object-level and point-level results as well as variation across rotations.

  2. Clarification and analysis request: Can you quantify frame consistency across normal instances, the change in the predicted frame between a normal sample and its defective counterpart, and performance on categories with strong rotational symmetries? This would help explain why the learned canonicalizer is substantially better than random equivariant weights on some metrics but differs by only 0.1 P-AUROC on Real3D-AD.

  3. Clarification: How does the canonicalizer differ in practice from prior self-supervised point-cloud canonicalization methods such as ConDor? Please also describe the exact sampling and replacement rule for the category-wise canonical queue. These details would help assess the specific novelty and reproduce the prototype loss.

Reviewer_2YeZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a rotation-invariant method for detecting defects in 3D point clouds across multiple categories. An SO(3)-equivariant canonicalizer aligns inputs to a consistent orientation, using a pairing loss to maintain alignment under synthetic defects and a prototype loss to encourage within-category consistency. The canonicalizer is then frozen while a DGCNN feature extractor, geometry-aware local attention module, and point-wise scoring head are trained on pseudo-anomalies added to normal objects. Experiments on Real3D-AD and Anomaly-ShapeNet show strong point-level AUROC.

Critical strengths
  1. The method handles multiple object categories and arbitrary input poses with one model, which is computationally efficient.
  2. The method achieves good empirical results on point cloud anomaly detection benchmarks.
Critical weaknesses
  1. The novelty is incremental, as the framework combines established components. Aligning inputs before anomaly detection is not a new concept: Reg3D-AD [1] already registers each test point cloud to a normal reference. The proposed SO(3)-equivariant canonicalizer offers an alternative way to achieve alignment.

  2. The method relies on pseudo-anomaly training, but neither discusses nor compares against closely related work such as PO3AD [2] and HPPF [3].

[1] Liu et al. Real3D-AD: A Dataset of Point Cloud Anomaly Detection. NeurIPS 2023 Datasets & Benchmarks Track.

[2] Ye et al. PO3AD: Predicting Point Offsets toward Better 3D Point Cloud Anomaly Detection. CVPR 2025.

[3] Kang et al. Hierarchical Point-Patch Fusion with Adaptive Patch Codebook for 3D Shape Anomaly Detection. CVPR 2026.

Questions for authors
  1. How does the proposed pseudo-anomaly training strategy differ from those in PO3AD [2] and HPPF [3]? Could the authors discuss these closely related methods and evaluate them under a comparable setting?

  2. On Real3D-AD, the random and trained canonicalizers achieve nearly identical point-level performance (Table 2). Does this suggest that the equivariant architecture alone is sufficient for localization on this dataset? What additional benefit do the proposed canonicalizer training losses provide for localization?

Reviewer_dSp1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CRU3D-AD, a unified 3D anomaly detection framework designed to handle multiple object categories and arbitrary input poses. The method employs an equivariant canonicalizer to learn canonical representations, with a pairing loss and a prototype loss to ensure defect robustness and class consistency. Furthermore, it combines geometry-aware local attention with synthetic anomaly generation to enable accurate point-wise anomaly scoring. Experiments on Real3D-AD and Anomaly-ShapeNet demonstrate superior point-level localization performance in the unified setting.

Critical strengths
  1. The paper introduces two constraints tailored to 3D anomaly detection: defect-robust and class-consistent alignment. The pairing loss and prototype loss are designed to maintain a stable canonical frame in the presence of defects and ensure consistency across samples of the same category.
  2. The paper provides extensive experimental evaluations, including robustness to test-set rotation and ablation studies, which support the effectiveness of the proposed CRU3D-AD framework.
Critical weaknesses
  1. The loss function ablation study in Table F should be revised to evaluate different combinations of loss terms, similar to the experimental design in Table 2, to provide a more comprehensive assessment of their effectiveness.
  2. Although the paper provides qualitative comparisons, a more systematic analysis of representative failure cases is needed to better understand the limitations of CRU3D-AD and provide further insights for future research.
Questions for authors
  1. Table F shows that removing the pairing loss improves P-AUROC on Real3D-AD from 85.5% to 86.7%. How do the authors explain this observation?
  2. In Table 2, the randomly initialized already achieves competitive performance on P-AUROC in P-AUROC. Could the authors explain why random canonicalization is so effective and further clarify the benefits of the learned canonicalizer?;
  3. Could the authors provide and analyze representative failure cases of CRU3D-AD? Such an analysis would help readers better understand the limitations of the proposed framework and provide valuable insights for future research.

Comments on the paper

1 comment

Dear Reviewers,

A friendly reminder to check the authors’ rebuttals and engage in the discussion, as the deadline is approaching. Authors still have time to address any further questions or concerns you may have, so your timely feedback would be greatly appreciated.

Thank you for your time and effort!

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.