Anomaly Factory 3D Diverse Pseudo-Anomaly Synthesis for Unsupervised 3D Anomaly Detection

Ali Balapour^1,2,3, Faraz Hach^1,2,3,*

Accepted to ECCV 2026

arXiv Code

TL;DR: AF3AD is a pseudo-anomaly generator, not an anomaly detector: it manufactures diverse, controllable geometric defects on normal point clouds and returns synthetic anomalous samples plus masks for downstream training.

Abstract

Detecting and localizing defects in 3D point clouds is challenging because abnormal samples are scarce and diverse, while training is often limited to normal data. We propose Anomaly Factory 3D (AF3AD), a modular framework that synthesizes diverse pseudo-anomalies from normal point clouds to expand the training data for unsupervised 3D anomaly detection methods that rely on pseudo-anomalies. AF3AD uses a center-conditioned parametric deformation model defined in local PCA frames, with kernel-controlled spatial falloff, anisotropy, directional gating, and normal/tangential displacement fields, enabling a broad set of geometric defect presets. We demonstrate its ease of use and effectiveness by integrating AF3AD with an offset-prediction detector and a reconstruction-based anomaly detection method, showing that AF3AD transfers across detection paradigms. Experiments on AnomalyShapeNet and Real3D-AD show consistent improvements in object- and point-level detection and localization, supported by ablations on preset groups and robustness under noise. AF3AD is designed as a standalone synthesis tool to facilitate adoption across different 3D anomaly detection paradigms.

InputNormal-only point clouds

Generator11 controllable geometric presets

OutputPseudo-anomalies and GT masks

IntegrationOffset and reconstruction models

AF3AD pipeline showing point cloud inputs, parameter sampling, PCA, point-wise deformation, outputs, anomaly detector integrations, and anomaly preset groups. — Pipeline overview: AF3AD samples local frames and deformation parameters from normal point clouds, then synthesizes pseudo-anomalous geometry with point-wise masks.

How it works

From a normal point cloud, AF3AD samples a local segment, builds a PCA frame at the anomaly center, shapes a smooth falloff with a spatial kernel, and displaces each point along a selected normal or tangent direction.

Choose a local frame

Partition the cloud into 64 patches, sample a center, and estimate tangent axes u, v and normal n using local PCA.

Shape the affected region

Use kernel falloff, anisotropy, soft gating, radius, and magnitude to define a plausible defect footprint.

Deform and supervise

Move points along normal or tangential fields while retaining the generated point-wise anomaly mask.

p′ = p + s · B · K(t) · g · d

The presets in the demo are curated points in this parameter space; free explore exposes the raw knobs directly. In the synthesis module, magnitude and radius are sampled from bounded Beta distributions, giving each preset controlled stochastic variation.

Interactive preview

Explore the pseudo-anomaly factory

Try the preset library or tune the deformation knobs directly. The browser demo is simplified, but it mirrors the core idea: controllable geometry changes with a corresponding anomaly mask.

n u v high displacement

drag to rotate

Radius R0.30

Magnitude B0.16

Configuration

Basic bulge

Points affected

–

Synthesis library

Eleven geometric defects under one formulation

AF3AD unifies isotropic, anisotropic, asymmetric, periodic, and tangential surface changes in a single center-conditioned deformation model. The output is a library of generated pseudo-anomalies and masks, giving downstream detectors a broader set of abnormal cues than generators limited to only bulges and dents.

Animated 4 by 3 gallery of AF3AD-generated pseudo-anomaly examples across multiple object categories. — Preset gallery on Real3D-AD samples. The AF3AD pseudo-anomalies in this GIF are intentionally exaggerated for visualization, so they should be read as qualitative examples rather than real defect severity.

ANormal single-lobe

Basic bulges and dents: the simple normal-direction deformations that match common pseudo-anomaly baselines.

BStructured marks

Elongated ridges, trenches, flat spots, and skewed impact craters with anisotropic support and directional gating.

CCompound patterns

Double-sided ripples and micro-dimple fields that expose the detector to repeated or multi-frequency geometry changes.

DTangential shifts

Shear, slip, and directional drag presets that move points along tangent directions instead of only along normals.

Design intent: AF3AD is modular by construction. New presets can be added by choosing a kernel, anisotropy, sign, gating rule, and displacement direction within the same shared deformation formulation.

Results

Downstream detectors benefit from richer pseudo-anomalies

Across AnomalyShapeNet and Real3D-AD, detectors trained with AF3AD-generated samples improve object-level detection and point-level localization over strong pseudo-anomaly baselines.

AnomalyShapeNet91.5O-AUROC, +5.5 over the previous best reported method.

AnomalyShapeNet92.5P-AUROC for point-level anomaly localization.

Real3D-AD85.2O-AUROC, +7.1 over the previous best reported method.

Generator cost0.96msMean synthesis cost per call when invoked by a downstream dataloader.

Comparison with strong baselines

Dataset	Method	O-AUROC	P-AUROC
AnomalyShapeNet	PO3AD	83.9	89.9
AnomalyShapeNet	Reg2Inv	86.0	88.2
AnomalyShapeNet	PO3AD trained with AF3AD samples	91.5 ± 2.1	92.5 ± 1.6
Real3D-AD	PO3AD	76.6	—
Real3D-AD	Reg2Inv	78.1	87.9
Real3D-AD	PO3AD trained with AF3AD samples	85.2 ± 1.2	86.1 ± 1.9

Dash indicates that the metric was not reported for that baseline in the manuscript comparison table.

Modularity with reconstruction-based detection

Setting	Downstream detector	O-AUROC	P-AUROC	Takeaway
AnomalyShapeNet, 15 classes	R3D-AD trained with AF3AD samples	67.2	60.7	+4.5 O-AUROC over the Patch-Gen instantiation.

Qualitative anomaly localization results comparing ground-truth masks and predicted anomaly score maps on Real3D-AD and AnomalyShapeNet samples. — Qualitative downstream localization examples. Red regions show ground-truth anomaly masks; yellow-white heat maps show predictions from a detector trained using AF3AD-generated pseudo-anomalies.

Ablations

Diversity is the useful signal

The manuscript separates detector architecture changes from synthesis changes. Under the matched multi-head offset predictor with hidden size 128, Norm-AS reaches 91.6 / 88.9 O/P-AUROC on the 15-class AnomalyShapeNet split, while the same detector trained with AF3AD-generated samples reaches 97.0 / 93.6.

Preset groups compound

Combining groups A-D outperforms single-group synthesis, showing that downstream detectors benefit from broader generated defect coverage.

Noise robustness holds

Mild Gaussian jitter leaves performance essentially unchanged; stronger noise mainly affects fine point-level precision.

Representative UMAP feature-space visualization for Real3D-AD airplane anomaly regions showing real test anomalies in blue, AF3AD synthesis in green, and Norm-AS synthesis in red. — Representative Real3D-AD feature-space UMAP. Blue denotes real test anomalies, green denotes AF3AD synthesis, and red denotes Norm-AS. The metrics below are mean values over 12 categories; this is offline alignment evidence using test-anomaly features as a proxy, not deployment-time performance.

Coverage ↑0.380 vs 0.087AF3AD covers more real-anomaly feature neighborhoods than Norm-AS.

MMD ↓0.105 vs 0.326Lower distribution distance in the original feature space.

Hull IoU ↑0.483 vs 0.157Higher overlap in the UMAP 2D projection.

Scope

What AF3AD models

AF3AD is designed for geometric surface defects in point clouds: bulges, dents, ridges, trenches, shear, drags, flat spots, and related smooth deformations. It directly modifies geometry and therefore does not model purely appearance-based or material defects. It also does not perform anomaly detection by itself; detection is handled by whichever downstream model consumes its generated samples.

Sharp cracks, scratches, and very high-frequency damage are only partially covered by the current smooth-kernel presets. The manuscript treats this as a limitation and leaves dataset-guided preset selection and richer generative synthesis as future work.

Citation

@misc{balapour2026anomalyfactory3dmodular,
      title={Anomaly Factory 3D: A Modular Framework for Diverse Pseudo-Anomaly Synthesis in Unsupervised 3D Anomaly Detection}, 
      author={Ali Balapour and Faraz Hach},
      year={2026},
      eprint={2606.29181},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.29181}, 
}