PA3AD for 3D Anomaly Detection

• PA3AD is a physics-inspired framework that generates pseudo anomalies by mimicking industrial defect morphologies (bulge, concavity, hole, crack, and bend) from normal 3D point clouds.
• The method employs a prototype-guided, shared-weight backbone with local-global attention and a difference-aware offset regression head to learn geometric deviations directly.
• PA3AD overcomes reconstruction pitfalls by explicitly combining synthetic defect generation and offset prediction, achieving superior anomaly detection on datasets like Anomaly-ShapeNet and Real3D-AD.

Searching arXiv for PA3AD and closely related 3D anomaly-detection papers to ground the article in current literature. PA3AD, short for Physics-inspired Pseudo Anomaly Generation and Prototype Feature Guidance for 3D Anomaly Detection, is a supervised-by-synthesis framework for point-cloud anomaly detection and localization under one-class training, where only normal point clouds are available during training. It addresses the scarcity and acquisition cost of real anomalous 3D samples by generating physically plausible pseudo anomalies from normal data and combining them with a prototype-guided, weight-sharing backbone and a difference-aware offset regression head. The method is designed for industrial 3D inspection settings in which realistic defect morphology matters and arbitrary perturbations can bias the detector away from true anomaly structure (Ning et al., 12 Jul 2026).

1. Problem setting and conceptual basis

PA3AD is formulated for one-class learning on 3D point clouds. The training set contains only normal samples,

$$D_{\text{train}}=\{P_i\in \mathbb{R}^{N\times 3}\}_{i=1}^{M},$$

while the test set contains labeled normal and anomalous point clouds,

$$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$

The detector is required to output point-wise anomaly scores,

$$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$

from which an object-level score is derived (Ning et al., 12 Jul 2026).

The central motivation is twofold. First, real anomalies are scarce and expensive to capture in industrial inspection data, so anomaly-free training is the typical regime. Second, pseudo-anomaly generation should not be reduced to random deletion, jitter, or unconstrained local perturbation. PA3AD argues that synthetic supervision is more useful when it follows the morphological signatures of industrial failures, including bulges, concavities, holes, cracks, and bends. In this sense, the framework treats anomaly synthesis not as data augmentation in the ordinary sense, but as an explicit approximation to defect formation.

The method also addresses a known weakness of reconstruction-based approaches: if the model is trained only to reconstruct normal geometry, anomalous inputs may be projected back toward the normal manifold or may trigger an identity-shortcut regime. PA3AD instead supervises the detector with synthetic defects and asks it to learn geometric deviation directly through point-wise offset regression. This suggests a shift from pure normality modeling toward explicit abnormality modeling, while still remaining within a one-class training setup.

2. System architecture

PA3AD has four main components: an MP-AG pseudo-anomaly generator, a shared-weight backbone with local-global attention, a momentum-based normal prototype, and a difference-aware fusion and offset prediction head (Ning et al., 12 Jul 2026).

The architecture operates on paired inputs during training: an original normal point cloud and a pseudo-anomalous counterpart derived from it. Both are passed through the same backbone so that normal and anomalous representations inhabit a common feature space. This shared-weight design is important because the later fusion and discrepancy modeling assume that differences between the two streams are semantically meaningful rather than artifacts of separate encoders.

The framework then maintains a normal prototype as a stable latent reference and uses a fusion module that explicitly amplifies discrepancies between normal and pseudo-anomalous features. The final prediction head regresses point-wise offsets, with the intended interpretation that anomalous regions correspond to geometric displacements from a normal surface.

At inference time, pseudo-anomaly generation is disabled. A test cloud is encoded, compared against the learned normal prototype, and passed through the same discrepancy-aware pipeline to produce point-wise offsets. The anomaly score is then derived from offset magnitude rather than from a reconstruction residual or a nearest-neighbor retrieval criterion. A plausible implication is that PA3AD attempts to unify localization and interpretability: if an anomaly is represented as a displacement field, then both its presence and its local geometric character are encoded in the same output.

3. Physics-inspired pseudo-anomaly generation

The pseudo-anomaly generator is termed MP-AG and models five defect morphologies: bulge, concavity, hole, crack, and bend (Ning et al., 12 Jul 2026). Its design principle is to simulate spatial deformation patterns associated with industrial defects rather than inject generic noise.

Let

$P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$

be a normal point cloud. For bulges and concavities, the deformation is defined along the surface normal $n_i$ around a center $c$ by a Gaussian radial basis function:

$$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$

Here, $\alpha>0$ corresponds to a bulge and $\alpha<0$ to a concavity, while $\sigma_{bh}$ controls spatial spread. This yields a smooth local protrusion or depression with radial decay.

For hole defects, the method uses a hyperbolic tangent profile,

$$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$0

and removes points satisfying

$$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$1

The result is a missing center with a depressed boundary, intended to mimic material loss.

For cracks, the method approximates crack orientation by PCA on the $$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$2-nearest neighbors around the crack center $$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$3. Let $$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$4 be the principal eigenvector. Points are then displaced in opposite directions according to their position relative to the crack plane:

$$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$5

This produces a slit-like separation along an estimated dominant direction.

For bends, the deformation is inspired by Euler–Bernoulli beam deflection:

$$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$6

with

$$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$7

where $$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$8 is the bending strength, $$D_{\text{test}}=\{(P_k,t_k)\}_{k=1}^{K}, \qquad t_k\in\{0,1\}.$$9 the bending axis, $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$0 the vector from $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$1 to the bending axis, $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$2 a smoothing weight, and $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$3 a numerical constant.

During training, pseudo-anomaly types are sampled with probabilities 0.35 for bulge, 0.35 for concavity, 0.10 for hole, 0.10 for crack, and 0.10 for bend. The deformation strength parameters are sampled from the default range $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$4, and the default scaling is reported to achieve the best performance. The generated anomalies are described as preserving continuity of surface deformation, local geometric plausibility, approximate defect morphology, and spatial coherence with the object surface. This suggests that the generator is intended not merely to diversify training data, but to impose an inductive bias about what a plausible 3D anomaly should look like.

4. Backbone, prototype guidance, and discrepancy modeling

The point cloud is first voxelized as

$$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$5

where $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$6 is voxel size, $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$7 denotes coordinates of non-empty voxels, and $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$8 aggregated voxel features. The reported voxel size is $$\phi:\mathbb{R}^{N\times 3}\rightarrow \mathbb{R}^{N},$$9 (Ning et al., 12 Jul 2026). The backbone is an encoder-decoder with sparse convolutions and a local-global attention module at the bottleneck.

The local branch partitions the feature matrix $P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$0 into overlapping cubic windows of side length $P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$1 and stride $P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$2. Relative positional encoding is derived from pairwise coordinates:

$P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$3

Multi-head self-attention is then applied within each window. The role of this branch is to capture local defects such as edges, cracks, and small deformations.

The global branch uses linear attention:

$P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$4

$P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$5

where $P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$6, $P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$7 is ReLU, and $P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$8 is a normalization diagonal matrix. The purpose of the global branch is to encode long-range structure, symmetry, and broader shape context. The local and global branches are fused into the final backbone representation.

PA3AD also maintains a momentum-based normal prototype. Let $P=\{p_i\in\mathbb{R}^3\}_{i=1}^N$9 be the prototype at iteration $n_i$0, and let $n_i$1 be the batch-mean normal feature. The prototype is updated by exponential moving average:

$n_i$2

The paper further expands this as

$n_i$3

The momentum $n_i$4 increases from 0.1 to 0.999 during training. The authors interpret this as a low-pass filter with effective window size roughly $n_i$5. In practice, the converged prototype is stored and used directly at inference. The stated rationale is that the prototype serves as a normality anchor that stabilizes representation learning and mitigates overfitting to synthetic artifacts.

After backbone encoding, the normal and pseudo-anomalous streams produce

$n_i$6

These are concatenated,

$n_i$7

compressed by an MLP,

$n_i$8

and modulated using the raw feature difference

$n_i$9

The difference branch produces adaptive weights,

$c$0

which reweight the fused representation elementwise:

$c$1

This difference-aware fusion block explicitly emphasizes latent dimensions where normal and pseudo-anomalous features diverge.

5. Offset learning, objective functions, and inference

The enhanced representation is mapped to point-wise offsets

$c$2

through an MLP-based head (Ning et al., 12 Jul 2026). The model is therefore trained to explain anomaly regions as geometric displacements relative to normal structure.

The total loss is

$c$3

with $c$4 and $c$5.

The offset norm loss is

$c$6

where $c$7 is the ground-truth offset induced by the pseudo-anomaly generator and equals zero for normal points. This term aligns predicted and target displacement magnitudes.

The offset direction loss is

$c$8

with $c$9. It encourages agreement in offset direction.

The feature discrepancy loss is

$$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$0

where $$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$1 and $$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$2 denote normal and anomalous features. If the two become too similar, the penalty grows. The effect is to maintain separation between normal and pseudo-anomalous representations.

The weight regularization term is

$$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$3

with $$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$4. This prevents the adaptive weights from collapsing to extreme values.

Training jointly uses the original normal cloud, the synthesized pseudo-anomalous cloud, and the corresponding synthetic offset labels. The detector learns to produce nonzero offsets for pseudo anomalies and near-zero offsets for normal data. At inference, the pseudo-anomaly generator is not used. A test cloud is encoded, compared with the learned normal prototype, and passed through the difference-aware pipeline. The point-level anomaly score is

$$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$5

and the object-level score is

$$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$6

Higher scores indicate greater anomaly likelihood.

6. Empirical results, ablations, and limitations

PA3AD is evaluated on Anomaly-ShapeNet, Real3D-AD, and supplementary MVTec 3D-AD results (Ning et al., 12 Jul 2026). Anomaly-ShapeNet contains 40 categories and 1,600 samples, with 4 normal training samples per category. Real3D-AD contains 12 categories of real industrial objects, with 4 normal training samples and 100 test instances per category.

On Anomaly-ShapeNet, the reported metrics are O-AUROC, P-AUROC, and O-AUPR. The paper states that PA3AD surpasses prior methods by 8.4% in object-level AUROC and achieves a mean O-AUPR of 95.2%, about 8% higher than existing methods. It is described as consistently top-performing across categories and attains several 1.000 results. Relative to methods including BTF, M3DM, PatchCore variants, IMRNet, R3D-AD, PO3AD, MC3D-AD, Reg2Inv, and Simple3D, it is usually best or among the best in both object-level and point-level metrics.

On Real3D-AD, PA3AD achieves strong average object-level and point-level performance and is reported to be particularly effective on categories with long-range or curved geometry, such as fish and seahorse. The stated explanation is the combination of local-global attention and physically realistic pseudo anomalies. The paper also notes that categories such as airplane remain difficult because Real3D-AD is challenging: scans are single-sided, dense and sparse sampling is highly non-uniform, and anomalies may occur in occluded regions.

In the appendix, the method is also reported on MVTec 3D-AD, achieving the best mean O-AUROC of 93.3%, compared with 91.8% for the second-best CPMF, while reaching a mean P-AUROC of 96.5%.

The ablation studies are central to the paper’s argument. For pseudo-anomaly generation, MP-AG is compared with random point deletion, random displacement, and Gaussian noise. On selected categories, the reported scores are substantially higher for MP-AG: AnomalyShapeNet bucket0: 0.968 vs 0.657 / 0.711 / 0.783, Real3D car: 0.769 vs 0.468 / 0.402 / 0.531, and Real3D seahorse: 0.875 vs 0.527 / 0.473 / 0.614. This supports the claim that defect-shaped, continuous deformations are more useful than generic perturbations.

For prototype guidance, three variants are compared: $$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$7 with pseudo anomalies only, $$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$8 with prototype-guided normal features, and $$p_i' = p_i + \alpha \cdot \exp\!\left(-\frac{\|p_i-c\|^2}{2\sigma_{bh}^2}\right) n_i.$$9 with prototype-guided features plus momentum update. Performance improves monotonically: AnomalyShapeNet O-AUROC: 0.874 $\alpha>0$0 0.921 $\alpha>0$1 0.936 and Real3D O-AUROC: 0.724 $\alpha>0$2 0.768 $\alpha>0$3 0.789. The momentum schedule is also ablated: fixed $\alpha>0$4 is described as noisy and unstable, fixed $\alpha>0$5 as too sluggish early in training, cosine 0.1 $\alpha>0$6 0.999 as strong, and linear 0.1 $\alpha>0$7 0.999 as best.

For the attention mechanism, the combined local + global design achieves the highest AUROC, with local attention contributing more strongly and global attention providing complementary robustness. For the loss function, the best results are obtained by using all four terms; norm + dir is already strong, feature discrepancy yields the largest additional gain, and weight regularization provides further stabilization. The best hyperparameter balance is reported for $\alpha>0$8 and $\alpha>0$9, while stronger weight regularization degrades performance.

The paper reports that the model can process full-resolution point clouds of about 44K points with inference time around 13.13 ms and GPU memory 0.36 GB on an RTX 4090, although it also notes that the backbone has a large parameter count. Its stated limitations include difficulty on some Real3D-AD categories under viewpoint bias and sparse sampling, incomplete coverage of all possible industrial defect morphologies by the five-type generator, open domain-shift problems across acquisition settings, and the relatively large parameter count of the backbone.

Taken together, these results position PA3AD as an end-to-end 3D anomaly detection framework whose distinctive contribution lies in aligning three mechanisms: physics-inspired pseudo-defect synthesis, stable prototype-based normality anchoring, and explicit discrepancy-driven offset regression. This suggests that its main significance is methodological rather than merely empirical: it replaces arbitrary synthetic corruption with morphology-aware supervision and uses that supervision to train a detector that models anomaly as structured geometric deviation rather than as residual error alone (Ning et al., 12 Jul 2026).