IEC3D-AD: Industrial 3D Anomaly Detection Dataset

• IEC3D-AD is a high-fidelity dataset featuring dense, point-level annotations that enable precise localization of machining-induced defects.
• It comprises 2400 3D scans across 15 industrial equipment categories, captured with a multi-sensor system for enhanced data quality.
• The accompanying GMANet paradigm uses synthetic point generation and spatial discrepancy optimization to achieve state-of-the-art anomaly detection.

The Point Cloud Anomaly Detection Dataset IEC3D-AD is a high-fidelity, point-level annotated benchmark for unsupervised 3D anomaly detection in authentic industrial environments. Designed to address the limitations of prior datasets such as MVTec 3D-AD and Real3D-AD, IEC3D-AD provides dense and granular 3D scans of 15 categories of real industrial equipment components (IECs), acquired directly from production lines. The dataset is accompanied by a novel training paradigm, GMANet, specifically tailored to leverage point-level spatial discrepancies for enhanced anomaly detection performance in demanding manufacturing scenarios (Guo et al., 5 Nov 2025).

1. Dataset Composition and Acquisition

IEC3D-AD comprises 2400 point cloud samples spanning 15 distinct industrial equipment component categories, with each category contributing 100 normal and 60 abnormal samples:

Category	Normal Samples	Abnormal Samples
Fender Ring	100	60
Washer	100	60
Butterfly Nut	100	60
Butterfly Bolt	100	60
Jointing Stud	100	60
Slipknot Bolt	100	60
K Nut	100	60
Hole Retaining Ring	100	60
Hexagonal Nut	100	60
Hexagonal Bolt	100	60
Pulling Nut	100	60
Double End Stud	100	60
Square Welded Nut	100	60
T Screw	100	60
Round Nut	100	60

Scans are acquired using a multi-sensor system integrating a DLP 3010 structured-light projector (1280×720), dual Daheng MER-502-79U3M CCD cameras (2/3″ IMX250 global-shutter CMOS), and a 360° automatic turntable. The resulting point clouds range from approximately 14K to 503K points per sample, with a mean of 160K points. The acquisition pipeline sequentially applies intrinsic/extrinsic camera calibration (Zhang’s method), structured-light multi-view capture, coarse and fine registration (4PCS, ICP), followed by density-based outlier removal and octree-based downsampling.

Abnormal samples are annotated using CloudCompare, featuring point-level binary masks for machining-induced defects: convex protrusions, concave pits, scratches, scars, and deformations. The anomaly-point ratio per sample spans 0.78% to 2.28% (mean ≈1.35%), providing fine-grained supervision signals.

2. Comparative Analysis and Dataset Properties

IEC3D-AD distinguishes itself quantitatively and qualitatively from existing 3D anomaly detection datasets:

Dataset	Categories	Samples	Points/Sample	Total Points (×10³)	Sample Type
MVTec 3D-AD	10	3604	10K–20K	—	Real
Real3D-AD	12	1200	35K–780K	224,720	Real
Anomaly-ShapeNet	40	1600	8K–30K	—	Synthetic
IEC3D-AD	15	2400	14K–503K	340,335	Real

Relative to previous real-world datasets, IEC3D-AD provides a higher number of categories, an expanded total point count (≈340M vs. 225M for Real3D-AD), an order-of-magnitude denser sampling (mean 160K vs. 15K per sample for MVTec 3D-AD, i.e., ~10.7× density ratio), and crucially, point-level defect annotations. The dataset exclusively features genuine industrial components, in contrast to prior reliance on artificial objects or synthetic data.

Point-level defect masks surpass the object-level or coarse-masked annotations typical of alternative datasets, supporting precise localization and detailed evaluation of anomaly detection models. The dataset’s density and granularity directly support research on subtle and localized industrial defects that are difficult to capture or evaluate with lower-resolution data.

3. Defect Annotation and Granularity

Annotations in IEC3D-AD leverage CloudCompare for precise defect labeling at the point level, targeting traces of authentic production defects. Defect categories include convex protrusions, concave pits, scratches, scars, and deformations—representative of machining-induced errors. The annotation process yields fine-grained masks: typically, each scan contains between 0.78% and 2.28% labeled anomalous points, with a mean of 1.35%. This allows models to be evaluated both on their ability to detect the presence of defective samples (object-level) and to localize individual defects (point-level). The fidelity of these labels underpins both supervision and evaluation, facilitating rigorous benchmarking of anomaly localization capabilities.

4. GMANet: Unsupervised 3D-AD Paradigm

GMANet defines an unsupervised 3D anomaly detection paradigm, inspired by generative approaches in 2D-AD but tailored for point cloud data. The architecture consists of two principal modules:

• SPCG (Synthetic Point Cloud Generation): Synthetic anomalies are generated via geometric morphological analysis of normal samples. The process identifies high-curvature regions in the point cloud (via Poisson reconstruction), scores clusters, and augments the normal data by adding/removing a small fraction $\epsilon$ of points in the highest-score region. This results in synthesized point clouds that simulate plausible industrial defects.
• SDO (Spatial Discrepancy Optimization): Two parallel PointNet++ encoders (an "expert" with frozen weights $\Phi_e$ and an adaptive "apprentice" $\Phi_a$) extract features. Masked features for normal ($F_n$) and abnormal ($F_a$) regions compute L2 discrepancies, with focal-style weighting to emphasize hard-to-classify points. The loss function $\mathcal{L}_{\rm SDO}$ simultaneously minimizes the margin in the normal domain and maximizes it in the abnormal domain, promoting overlap and reducing separability in feature space.

Formally, features are extracted as:

$$F^e = \Phi_e(d;\theta_e),\quad F^a = \Phi_a(d;\theta_a)$$

and the final anomaly score for test data $d^-$ is evaluated as:

$$\mathcal{A}(d^-) = \bigl\|\Phi_e(d^-;\theta_e)-\Phi_a(d^-;\theta_a)\bigr\|_2.$$

Ablation analyses demonstrate both SPCG and SDO modules are critical for state-of-the-art performance.

5. Experimental Evaluation and Benchmarking

IEC3D-AD serves as both a data benchmark and a proving ground for unsupervised 3D-AD methodologies. Evaluation occurs at both the object and point levels, using AUROC and AUPR as metrics.

On IEC3D-AD (mean over 15 categories):

• Object Level: AUROC = 0.8803, AUPR = 0.8612
• Point Level: AUROC = 0.7252, AUPR = 0.1089

These metrics surpass baselines such as BTF, M3DM, PatchCore, Reg3D-AD, IMRNet, and R3D-AD by large margins. The ablation paper quantifies individual module contributions: models without SPCG drop to AUROC = 0.6578, AUPR = 0.0723; without SDO, AUROC = 0.6754, AUPR = 0.0832; compared to the full system’s AUROC = 0.7252 and AUPR = 0.1089.

Generalization experiments demonstrate transferability: on Real3D-AD, GMANet achieves AUROC = 0.741 (+5.3% over baseline), AUPR = 0.756 (+4.6%) at the object level, with comparable performance gains at the point level. Performance on Anomaly-ShapeNet modestly exceeds previous results (AUROC ≈ 0.673, AUPR ≈ 0.630), indicating robustness to object variability and defect heterogeneity.

Qualitative analyses demonstrate the consistent detection of microscopic edge defects and a reduced tendency toward over-segmentation in normal regions.

6. Significance and Implications for 3D Anomaly Detection

IEC3D-AD constitutes a new standard for industrial 3D anomaly detection research. Its high-resolution, point-dense data and fine-grained annotation minimize the abstraction gap between dataset and production environments. The design supports evaluation of both class-level and localization ability in unsupervised settings, promoting advances in anomaly interpretation and intervention in safety-critical industrial contexts. The dataset’s balance of authenticity, scale, and annotation detail allows for the development of more sensitive and robust defect detection methods. The accompanying GMANet paradigm demonstrates the practical benefit of synthesizing anomalies and leveraging spatial feature overlaps, offering an extensible methodology for other industrial and scientific domains focusing on fine-grained point cloud analysis.

A plausible implication is that future benchmark datasets in 3D-AD may require point-level labeling and authentic industrial diversity to facilitate further algorithmic progress. IEC3D-AD, in combination with methodologically aligned frameworks such as GMANet, is positioned to serve as both a benchmark and reference architecture for advanced research in unsupervised 3D point cloud anomaly detection.