Anomaly Factory 3D (AF3AD)

A modular factory for controlled geometric pseudo-anomalies in 3D point clouds.

Paper | Quickstart | Citation

AF3AD turns normal 3D point clouds into realistic pseudo-anomalous examples. It is built for unsupervised 3D anomaly detection, where real defective training samples are scarce but normal samples are available.

The core package is intentionally small: it takes points, normals, an anomaly center, and a preset configuration, then returns a deformed point cloud. The detector-specific training code is kept as reference integrations.

AF3AD selects a local region, builds a PCA frame, samples deformation parameters, applies anisotropic falloff and gating, then produces synthetic anomalies for downstream detectors.

Why AF3AD

• It is detector-independent: the synthesis core is not tied to one training architecture.
• It creates diverse geometric defects rather than a single generic bump.
• It supports normal and tangential deformation directions.
• It uses local PCA frames, anisotropy, kernels, and one-sided gating for controllable shape changes.
• It can plug into online training workflows, offline data generation, or custom anomaly-detection experiments.

At A Glance

Item	Details
Core dependency	NumPy
Input	Point positions, point normals, anomaly center, preset config
Output	Deformed point positions
Presets	11 geometric pseudo-anomaly types
Reference integrations	PO3AD-style offset prediction and R3D-AD-style reconstruction
Paper benchmarks	AnomalyShapeNet and Real3D-AD

Preset Library

AF3AD defines 11 geometric pseudo-anomaly presets:

#	Preset	Key characteristics
0	Basic Bulge	Isotropic outward bump (alpha=+1).
1	Basic Dent	Isotropic inward cavity (alpha=-1).
2	Ridge	Anisotropic bulge elongated along one tangent direction.
3	Trench	Anisotropic dent elongated along one tangent direction.
4	Elliptic Patch / Flat Spot	Pressed region using per-point normals.
5	Skewed Impact Crater	Oblique one-sided dent with global gating.
6	Shear U	Tangential slip along the first tangent direction.
7	Shear V	Tangential slip along the second tangent direction.
8	Double-Sided Ripple	Cosine ripple with stochastic sign.
9	Micro Dimple Field	Small dent used repeatedly for pitting-style defects.
10	Directional Drag / Stretch	Anisotropic tangential drag deformation.

Example pseudo-anomaly outputs produced by the preset library.

Each preset returns a SmartAnomaly_Cfg dataclass specifying support radius, displacement magnitude, kernel, anisotropy, displacement direction, gating mode, and random seed behavior.

Method In One Equation

AF3AD applies localized parametric deformations to normal point clouds. Around a selected anomaly center, it estimates a local PCA frame, transforms nearby points into local coordinates, evaluates an anisotropic distance, applies a falloff kernel, optionally gates the deformation to one side of the surface, and displaces points along normal or tangential directions.

p_i' = p_i + s * B * K(t_i) * g_i * d_i

where s controls bulge versus dent direction, B is the displacement magnitude, K(t_i) is the spatial falloff kernel, g_i is the optional gate, and d_i is the deformation direction.

Radius and magnitude are sampled from bounded Beta distributions, giving controlled variety without extreme deformations.

Repository Layout

af3ad/                  Reusable synthesis core
integrations/po3ad/     PO3AD-style online synthesis reference integration
integrations/r3dad/     R3D-AD-style offline synthesis reference integration
scripts/                Root-level entry points
configs/                Example experiment configs
assets/                 README figures and qualitative examples

Environment

The AF3AD core pseudo-anomaly synthesizer only requires NumPy.

Create a Python virtual environment named af3ad:

python3 -m venv ~/.venvs/af3ad
source ~/.venvs/af3ad/bin/activate

python -m pip install --upgrade pip
python -m pip install numpy

From the repository root, make the local package importable:

export PYTHONPATH="$PWD:$PYTHONPATH"

Quick check:

python -c "from af3ad import PseudoAnomalySynthesizer; print('AF3AD is ready')"

To leave the environment:

deactivate

Basic Usage

import numpy as np

from af3ad import PseudoAnomalySynthesizer

rng = np.random.default_rng(42)

points = rng.standard_normal((2048, 3)).astype(np.float32)
normals = points / (np.linalg.norm(points, axis=1, keepdims=True) + 1e-8)
center = points[rng.integers(len(points))]

synth = PseudoAnomalySynthesizer()
cfg = synth.preset_factory.presets[0]()  # Basic Bulge

deformed_points = synth.generate(points, normals, center, cfg)
print(deformed_points.shape)

List available presets:

for idx, name in synth.list_presets():
    print(f"[{idx}] {name}")

Customize the radius and magnitude sampling ranges:

class MyArgs:
    R_low_bound = 0.03
    R_up_bound = 0.20
    R_alpha = 2.0
    R_beta = 5.0
    B_low_bound = 0.01
    B_up_bound = 0.10
    B_alpha = 2.0
    B_beta = 5.0


synth = PseudoAnomalySynthesizer(args=MyArgs())

Integration Environments

The detector integrations require their own training environments and heavier dependencies such as PyTorch, CUDA-specific packages, MinkowskiEngine, Open3D, FAISS, and point-cloud processing libraries.

For PO3AD-based training, follow the original PO3AD repository:

https://github.com/yjnanan/PO3AD

For R3D-AD-based training, follow the original R3D-AD repository:

https://github.com/zhouzheyuan/r3d-ad

This repository packages the AF3AD pseudo-anomaly synthesis core and keeps detector-specific code as reference integrations.

Reference Commands

After installing the appropriate detector environment, the reference launchers can be run from the repository root:

python scripts/train_po3ad_integration.py --dataset Real3D --dataset-base-dir data/Real3D --category airplane
python scripts/train_po3ad_integration.py --dataset AnomalyShapeNet --dataset-base-dir data/AnomalyShapeNet/dataset --category ashtray0
python scripts/train_r3dad_integration.py --dataset-path ./data/dataset/pcd --category ashtray0
python scripts/generate_offline_anomalies.py integrations/r3dad/configs/shapenet-ad/base.yaml

Results From The Paper

The paper evaluates AF3AD on AnomalyShapeNet and Real3D-AD. For the offset-prediction instantiation, reported as mean +/- standard deviation over three seeds:

Dataset	O-AUROC	P-AUROC	Notes
AnomalyShapeNet	91.5 +/- 2.1	92.5 +/- 1.6	+5.5 O-AUROC over the previous best reported method.
Real3D-AD	85.2 +/- 1.2	86.1 +/- 1.9	+7.1 O-AUROC over the previous best reported method.

Ground-truth anomaly regions and predicted anomaly score maps on Real3D-AD and AnomalyShapeNet.

Scope

AF3AD focuses on geometric surface deformations such as bulges, dents, ridges, trenches, shear, and drag. It is not intended to synthesize purely appearance-based anomalies, material defects, or all high-frequency crack and scratch patterns exactly. The framework is meant to provide controllable geometric pseudo-anomaly diversity for training anomaly detectors.

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},
}

Release TODO (Weighted)

• Release weights