Reg3D-AD: 3D Anomaly Detection Benchmark

• Reg3D-AD is a registration-based 3D anomaly detection algorithm that integrates rigid registration and multi-resolution feature matching to accurately detect defects in high-precision industrial parts.
• It constructs dual memory banks from pre-trained PointMAE features and point coordinates, enabling few-shot learning and efficient anomaly scoring on high-resolution point cloud data.
• The benchmark demonstrates competitive object-level accuracy while highlighting limitations in point-level precision and challenges with non-rigid deformations.

Reg3D-AD refers to a registration-based 3D anomaly detection algorithm designed for high-precision industrial inspection, with particular emphasis on point cloud data. It is established as a non-parametric, memory-bank-driven method integrating both rigid registration and multi-resolution feature matching for the detection and localization of surface and structural defects in manufactured objects. Reg3D-AD was introduced as a baseline and benchmark method alongside the release of the Real3D-AD dataset, which offers high-resolution, 360-degree point cloud representations of manufactured parts for anomaly detection research (Liu et al., 2023).

1. Problem Setting and Context

Reg3D-AD is situated in the context of industrial quality control, where point cloud data acquired via 3D scanners is analyzed to detect localized anomalies (defects) on the surfaces or in the structure of produced parts. Given the scarcity of both extensive datasets and robust baseline algorithms for this purpose, the Real3D-AD dataset was established, and Reg3D-AD was proposed as a strong "few-shot," full-3D, point-level anomaly detection method tailored for high-precision equipment where resolution and coverage are critical.

2. Algorithmic Pipeline

The Reg3D-AD anomaly detection pipeline is organized into two main stages: prototype indexing (training) and inference (detection):

• Prototype Indexing: A small set of “perfect” prototypes (usually ≤4 per category), each represented as a complete high-resolution 3D point cloud, is preprocessed. Local geometric features (raw XYZ coordinates) and global scene descriptors (via a pre-trained PointMAE transformer encoder) are extracted and stored in dual memory banks.
• Inference: Given an observed point cloud scan Q, the method performs:
  1. Rigid registration of Q to the most similar prototype P* using RANSAC and FPFH correspondences, obtaining optimal transformation parameters (R*, t*).
  2. Feature extraction for every aligned point in Q, generating local (coordinate) and global (patch embedding) features.
  3. Anomaly scoring by comparing each test feature to its nearest neighbors in the memory banks, combining local and global feature distances with an optional registration residual.

This pipeline is summarized in the following table:

Stage	Description	Components/Methods
Prototype Indexing	Build and compress memory banks from perfect parts	PointMAE, Coreset
Inference	Register test scan, extract features, compute anomaly scores	RANSAC+FPFH, Faiss, scorer

3. Key Algorithmic Components

Point Cloud Registration

Reg3D-AD employs classical RANSAC-based rigid alignment using FPFH (Fast Point Feature Histograms), which matches local geometric structures between the test scan and the prototype. The rigid transform (R*, t*) is estimated by minimizing the sum of squared Euclidean distances for matched inliers. The registration residual r_j for each test point quantifies its minimal alignment error relative to prototype correspondences.

Feature Memory Banks

Two memory banks are constructed:

• Local Memory Bank (M^l): Contains subsampled prototype point coordinates (ℝ³), yielding granular correspondence at high resolution.

• Global Memory Bank (M^g): Holds patch-level feature vectors (ℝ^768) extracted by passing spatial patches through a frozen PointMAE transformer (12-layer, 768-dim hidden). Both banks are compressed using greedy farthest point sampling (Coreset) to fixed size (default K=10,000).

Anomaly Scoring

Each point in the test scan is scored on the basis of its local and global features as follows:

1. Compute feature distances to nearest memory bank neighbors for both local and global features.
2. Apply PatchCore-style importance weighting to stabilize the scores in regions of variable density.
3. The anomaly score per point is:

$$s(q) = \alpha \cdot s^{reg} + \beta \cdot \frac{s^l + s^g}{2}$$

where $s^{reg}$ is the registration residual, with default weights $\alpha=0$, $\beta=1$; the registration term can be included if improvement is observed.

4. Implementation and Evaluation

Key architectural details include:

• Preprocessing of prototypes to ~1 million points; test clouds are typically downsampled to ~200,000 points.
• RANSAC+FPFH feature extraction, with inlier threshold τ~0.02mm.
• Memory banks managed via Faiss for fast nearest neighbor search, with b=5 used for PatchCore reweighting.

Reg3D-AD is non-parametric at test time, requiring no training aside from memory bank construction. All learning for feature extraction resides in the frozen PointMAE encoder.

In experimental evaluation on the Real3D-AD dataset:

• Object-level AUROC: Reg3D-AD (PointMAE) achieves 0.704, outperforming PatchCore (FPFH+xyz, 0.682), BTF (xyz, 0.635), PatchCore (PointMAE, 0.594), and other previous baselines.
• Object-level AUPR: 0.723 (next best: PatchCore FPFH+xyz, 0.676)
• Point-level AUROC/AUPR: 0.700/0.113 (compared to best PatchCore at 0.722/0.129)
• Category-specific AUROCs: Highest on Fish (0.915), Diamond (0.900), Chicken (0.852); lowest on Gemstone (0.417), Starfish (0.506).

Ablation reveals that increasing point group size from 8192 to 16384 or memory bank size from 10,000 to 20,000 produces minor variations in AUROC, indicating efficiency and robustness.

5. Comparative Position and Limitations

Reg3D-AD serves as a baseline for registration-based, memory-augmented, few-shot 3D anomaly detection, with a modular design that enables easy extension. While it achieves superior object-level accuracy, it trades a small amount of point-level precision. Reg3D-AD does not train a registration or feature network from scratch, relying entirely on pre-trained encoders and fixed memory banks. The model is especially suitable for application domains where high-quality prototypes are available, and rigid part geometry is assumed.

A notable limitation is the computational and storage overhead introduced by the memory banks, which can become significant for very large datasets, though Coreset sampling mitigates this for up to 20,000 memory entries. Additionally, performance can degrade if the registration step fails due to significant non-rigid deformation or poor initial alignment.

A plausible implication is that the reliance on rigid registration constrains Reg3D-AD to applications where such transformations are valid, and further, true anomalies coinciding with non-rigid or systematic geometric deviations may be underdetected.

6. Relevance in the Evolving 3D Anomaly Detection Landscape

Within the broader context established by the Real3D-AD dataset and other approaches like R3D-AD (diffusion-based reconstruction) (Zhou et al., 2024), Reg3D-AD typifies the classical paradigm of registration-followed-by-feature comparison. Ablation studies demonstrate its stability and competitive AUROC relative to reconstruction-based and memory bank-free methods; for example, R3D-AD achieves an AUROC of 0.734 (Real3D-AD) and 0.749 (Anomaly-ShapeNet) but at lower memory consumption and higher throughput by eschewing memory banks in favor of a diffusion-generator.

Reg3D-AD’s place in the field is as a baseline against which learning-based, generative, or more complex registration pipelines can be benchmarked. Its non-parametric, interpretable, and easily extensible design facilitates its adoption for both practical industrial settings and algorithmic evaluation.

7. Prospects and Future Work

Future enhancements may include:

• Integration of learned or deformable registration networks in lieu of RANSAC-based rigid fitting to accommodate a broader class of defects and object variations.
• Joint training of feature extractors specifically for anomaly detection, replacing frozen PointMAE with networks optimized on detection loss.
• Hybrid frameworks that combine the explicit memory bank approach of Reg3D-AD with generative or reconstructive modules (e.g., diffusion-based reconstruction as in R3D-AD), leveraging both prototypical alignment and strong modeling of typical and atypical variations.

In sum, Reg3D-AD defines a rigorous, transparent, and high-performing reference point for 3D anomaly detection leveraging registration and dual-resolution feature memory, with demonstrated efficacy on the Real3D-AD benchmark (Liu et al., 2023).