3D-ADAM: Industrial 3D Anomaly Benchmark
- The paper introduces 3D-ADAM, a benchmark leveraging real manufactured parts and multi-sensor RGB+3D data to enhance anomaly detection in industrial settings.
- 3D-ADAM is a comprehensive dataset featuring diverse defect classes and detailed machine element annotations, enabling feature-aware localization and robust defect identification.
- The dataset’s realistic capture conditions—variable lighting, poses, and sensor modalities—challenge current models, urging further advancements in industrial anomaly detection.
Searching arXiv for papers on 3D-ADAM and closely related 3D anomaly detection benchmarks/methods. 3D-ADAM, short for 3D Anomaly Detection in Advanced Manufacturing, is a large-scale, real-world, multi-sensor RGB+3D industrial anomaly detection dataset introduced to support high-precision anomaly detection and localization under conditions closer to industrial deployment than earlier benchmarks. It was designed to address limitations attributed to prior 3D anomaly datasets, especially limited defect diversity, limited machine-element coverage, overly controlled capture conditions, and insufficient realism for advanced manufacturing environments. The dataset combines real manufactured parts, real defect instances, multiple industrial depth sensors, multi-view capture in a working robotics cell, and machine-element-level annotations in addition to defect labels (McHard et al., 10 Jul 2025).
1. Concept and intended scope
3D-ADAM is positioned as a benchmark for 3D anomaly detection, RGB+3D anomaly detection, and localization in manufacturing settings where part pose, lighting, and visibility are not tightly controlled. Its stated objective is to be more representative of real industrial deployment, particularly for inspection scenarios in which parts are observed from multiple views with varying orientations, ambient illumination, and partial occlusions (McHard et al., 10 Jul 2025).
A central feature of the dataset is its explicit treatment of machine elements as annotation targets alongside defects. The paper identifies this as a key distinction from prior datasets because feature-level distinctions such as faces, edges, fillets, chamfers, holes, gears, and related structures matter in manufacturing inspection but are absent from earlier benchmarks. This suggests a shift from purely object-level anomaly detection toward feature-aware localization and classification, although the benchmark reported in the paper is framed around anomaly detection and localization rather than a full machine-element recognition task (McHard et al., 10 Jul 2025).
The dataset is also explicitly industrial in provenance. Unlike synthetic or heavily curated benchmarks, 3D-ADAM uses actual manufactured parts and actual physical defects, and it was captured in a real industrial environment rather than a highly controlled laboratory setting. The paper’s broader claim is that this realism exposes substantial weaknesses in models that perform strongly on simpler benchmarks (McHard et al., 10 Jul 2025).
2. Position within the 3D anomaly-detection literature
The paper situates 3D-ADAM against three reference datasets: MVTec 3D-AD, Eyecandies, and Real3D-AD (McHard et al., 10 Jul 2025). Among these, MVTec 3D-AD is described as the most widely used benchmark for unsupervised 3D anomaly detection and localization, comprising 4147 scans from 10 real-world object categories, with anomaly-free train/validation sets and a mixed test set containing 948 anomalous test images and 249 anomaly-free test images (Bergmann et al., 2021). That benchmark established the basic industrial-inspection setting in which a method is trained only on normal data and must produce both sample-level anomaly decisions and pixel-/point-level localization, but its capture conditions are comparatively controlled and its ontology is narrower (Bergmann et al., 2021).
Relative to MVTec 3D-AD, 3D-ADAM is described as more varied in defect classes, machine-element annotations, part categories, and capture conditions, and as being collected in-the-wild rather than in a controlled setup (McHard et al., 10 Jul 2025). Relative to Real3D-AD, the paper states that 3D-ADAM goes further in industrial realism and in the breadth of annotated defect and feature types, while including RGB + XYZ alignment from multiple sensors (McHard et al., 10 Jul 2025). Relative to Eyecandies, it emphasizes that Eyecandies uses synthetic parts and defects, whereas 3D-ADAM uses real manufactured parts and physical defects (McHard et al., 10 Jul 2025).
This positioning places 3D-ADAM within a broader transition in the field. Recent methods such as MC3D-AD target multi-category 3D anomaly detection with unified geometry-aware reconstruction (Cheng et al., 4 May 2025), DAS3D targets dual-modality RGB+depth anomaly detection with anomaly synthesis (Li et al., 2024), and CoGeoAD targets zero-shot 3D anomaly detection through hierarchical color-geometric fusion (Xu et al., 24 Jun 2026). The introduction of 3D-ADAM suggests that progress on such methods increasingly depends on datasets that better reflect real deployment conditions rather than benchmark regularities.
3. Dataset composition, ontology, and reported statistics
The paper reports the overall scale of 3D-ADAM in several places, and some of those figures differ across sections. The abstract reports 14,120 high-resolution scans, 217 unique parts, 4 industrial depth sensors, 27,346 annotated defect instances from 12 categories, and 8,110 annotations of machine element features, spanning relevant mechanical design form factors (McHard et al., 10 Jul 2025). The dataset section and table are described as giving 12,192 scans distributed across four sensors, while the breakdown table totals 14,120 images across the categories; the paper’s headline figure is therefore the larger 14,120 count (McHard et al., 10 Jul 2025).
A similar discrepancy is reported for category counts and machine-element annotations. The abstract says 28 categories, whereas the dataset section/table enumerates 29 distinct object categories (McHard et al., 10 Jul 2025). Likewise, the abstract reports 8,110 machine-element annotations, while the dataset section/table reports 27,346 machine-element annotations in parallel with 27,346 defect annotations, suggesting a one-to-one propagation of annotations per defect instance (McHard et al., 10 Jul 2025). These differences are part of the published description and are not resolved in the provided text.
The dataset includes 12 defect classes:
| Type | Classes |
|---|---|
| Defect classes | cuts, bulges, holes, gaps, burrs, cracks, scratches, marks, warping, roughness, over-extrusion, under-extrusion |
It also includes 16 machine element classes spanning:
| Type | Classes |
|---|---|
| Machine element classes | faces, edges, internal/external fillets, internal/external chamfers, holes, kerfs, tapers, indents, counterbores, countersinks, spur gear teeth, rack gear teeth, spiral gear teeth, clockwise helical gear teeth, counter-clockwise helical gear teeth |
The parts are described as forming a robot arm mechanism, based on the open-source BCN3D-Moveo system and augmented with additional parts from Thingi10K so that the full set of machine elements is represented (McHard et al., 10 Jul 2025). Listed examples include base, shoulder, forearm, elbow, wrist, gripper, and multiple gear subcomponents. The total table reports 217 parts, 14,120 images, 27,346 defect annotations, and 27,346 machine element annotations (McHard et al., 10 Jul 2025).
4. Acquisition environment and data representation
3D-ADAM was captured in a real industrial robotics applications development cell at HAL Robotics (McHard et al., 10 Jul 2025). The acquisition setup used four industrial 3D sensors:
| Sensor |
|---|
| MechMind LSR-L |
| MechMind Nano |
| Intel RealSense D455 |
| Stereolabs Zed 2i |
The mounting geometry was heterogeneous. An ABB IRB 120 robot arm positioned the MechMind Nano, RealSense, and Zed 2i, while the MechMind LSR-L was mounted above the cell (McHard et al., 10 Jul 2025). A part fixture allowed 360° rotation around the primary axis and 90° motion around a secondary axis. The view-acquisition protocol involved at least one full rotation of the primary axis, with scans every 45 degrees, and the secondary axis fixed at 0° and 30°, with additional orientations used when required for coverage (McHard et al., 10 Jul 2025).
Lighting was intentionally variable. Scanning occurred under indirect natural light, diffuse artificial light when needed, different times of day, and weather and seasonal variation (McHard et al., 10 Jul 2025). The paper presents this variability as a core design principle rather than a nuisance, since the goal is to approximate industrial deployment rather than laboratory repeatability.
For each scan, the dataset contains RGB images and XYZ point cloud data, stored as PNG RGB images and .PLY point clouds with a 1-to-1 pixel correspondence between RGB and XYZ data (McHard et al., 10 Jul 2025). The sensors are described as capturing six-channel images containing r, g, b, x, y, z (McHard et al., 10 Jul 2025). This representation places 3D-ADAM squarely in the multimodal RGB+3D regime now used by methods such as DAS3D and CoGeoAD, which explicitly rely on aligned appearance and geometry channels (Li et al., 2024).
5. Manufacturing process, defect creation, and annotation protocol
The parts were manufactured by Fused Deposition Modelling (FDM) in PLA using two Prusa i3, one Ender 3 Pro, and one Ultimaker 2 (McHard et al., 10 Jul 2025). Defects originated from two sources. Some defects occurred naturally during printing, with the paper noting that even some nominally “good” parts exhibited minor defects discovered during annotation; burrs and roughness often arose this way (McHard et al., 10 Jul 2025). Other defects were deliberately induced either by changing print parameters or by manual intervention. Process-induced defects included bulges, gaps, warping, over-extrusion, and under-extrusion, while handling- or manipulation-induced defects included cuts, holes, cracks, marks, and scratches (McHard et al., 10 Jul 2025).
The annotation pipeline is one of the dataset’s principal technical contributions. For each part instance, the dataset provides a part segmentation mask, a defect segmentation mask, and machine element bounding boxes with class labels (McHard et al., 10 Jul 2025). Foreground part masks were produced semi-automatically using Cutie and then manually checked and corrected. Defect and machine-element annotations were performed manually on MechMind Nano images because they provide the highest defect detail. A single expert annotator labeled all Nano images, and HAL Robotics experts provided verification and feedback (McHard et al., 10 Jul 2025).
Annotations were then propagated across sensors: homographic transforms were used to project Nano annotations onto images from the other sensors, after which the propagated labels were manually checked and corrected where needed (McHard et al., 10 Jul 2025). This procedure is notable because it aims to preserve label consistency across a multi-sensor RGB+3D benchmark while maintaining alignment between appearance and geometry. A plausible implication is that 3D-ADAM is intended not only for per-sensor evaluation but also for research on cross-sensor robustness.
6. Benchmark tasks, evaluation protocol, and empirical difficulty
The benchmark is reported in a supervised setting for 3D anomaly detection, RGB+3D anomaly detection, and localization (McHard et al., 10 Jul 2025). The split uses an anomaly-free training set containing all defect-free scans, while the defective scans are divided into validation and test with a 60:40 split between test and validation (McHard et al., 10 Jul 2025). The evaluated models are TransFusion, CPMF, 3DSR, and Shape-Guided, all of which are identified as methods known from MVTec 3D-AD benchmarks (McHard et al., 10 Jul 2025).
The reported metrics are standard: AUROC for image-level anomaly detection and AUPRO for pixel-level anomaly localization (McHard et al., 10 Jul 2025). This aligns 3D-ADAM with prevailing 3D-AD evaluation practice. For comparison, MVTec 3D-AD evaluates sample-level classification with AUROC and localization with [email protected], emphasizing that high false-positive rates can make segmentation practically meaningless when anomalies are small (Bergmann et al., 2021). 3D-ADAM’s use of AUROC and AUPRO preserves that industrial anomaly-detection framing.
The paper’s principal empirical conclusion is qualitative but unambiguous: all state-of-the-art models perform significantly worse on 3D-ADAM than on MVTec 3D-AD (McHard et al., 10 Jul 2025). The results table is said to appear in supplementary material rather than the main text, but the authors attribute the performance drop to the dataset’s broader defect ontology, increased geometry diversity, machine-element variation, sensor differences, variable lighting, pose changes, and partial occlusion (McHard et al., 10 Jul 2025).
This claim is consistent with broader trends in later 3D-AD research. For example, CoGeoAD reports strong zero-shot performance on MVTec3D-AD and Eyecandies but also treats cross-dataset transfer as a central challenge (Xu et al., 24 Jun 2026). Likewise, IEC3D-AD was introduced on the grounds that prior benchmarks do not capture the subtle defects and dense full-coverage point clouds required for industrial equipment inspection (Guo et al., 5 Nov 2025). Taken together, these developments suggest that benchmark difficulty is increasingly driven by realism rather than by the formal anomaly-detection task definition alone.
7. Industrial relevance, limitations, and research implications
The paper includes an external validation of annotation quality through an expert survey involving five industry experts from three manufacturing organizations spanning industrial R&D, digital process manufacturing, and robotics (McHard et al., 10 Jul 2025). Each expert labeled a random subset of defective scans equal to about 1% of the total dataset, receiving instructions and defect-free examples of the same parts but no annotated examples from the dataset itself, approximating a blind evaluation (McHard et al., 10 Jul 2025). Because the survey labels were 2D-based, the comparison used IoU and PRO. The reported results were mean IoU = 0.6, standard deviation = 0.04, mean PRO = 0.76, and standard deviation = 0.058 (McHard et al., 10 Jul 2025). The paper interprets this as strong agreement between expert labels and dataset ground truth, with experts tending toward slightly more false positives than false negatives.
The paper is also explicit about limitations. All parts are printed in PLA, limiting material diversity and surface-appearance variation. The dataset does not contain all possible industrial defect types. It does not provide complete closed 3D scans from all viewpoints, unlike some prior datasets; however, the authors argue that this is more realistic for industrial deployment, where full-view scans are often unavailable. The effect of changing material properties on defect visibility is not studied (McHard et al., 10 Jul 2025).
These limitations shape the dataset’s likely role in the literature. The paper argues that future methods will need to handle multi-view, multi-sensor, real-world noise, and that anomaly detection should become feature-aware, not only object-aware (McHard et al., 10 Jul 2025). It also suggests relevance to few-shot, zero-shot, and machine-element-aware localization and classification. This suggests that 3D-ADAM is not merely a larger benchmark, but part of a broader move toward deployment-oriented evaluation in 3D industrial inspection.
A common misconception would be to treat 3D-ADAM as simply a scaled-up version of earlier RGB+3D datasets. The published description indicates a more specific claim: the dataset’s novelty lies not only in scale, but in the combination of real industrial capture, multi-sensor aligned RGB+XYZ data, broad defect ontology, and machine-element annotations (McHard et al., 10 Jul 2025). Another possible misconception would be to assume that the dataset is internally uniform in all reported counts; the paper itself reports differing totals for scans, category counts, and machine-element annotations across sections, and those differences should be preserved when citing the benchmark rather than silently normalized (McHard et al., 10 Jul 2025).