MPW-AD: Multi-Photometric Weld Anomaly Detection Benchmark

Version: 1.0 Released with: MPF-AD (Multi-Photometric Fusion for Anomaly Detection), manuscript in preparation, 2026. License: CC-BY-4.0 Contact: https://github.com/WillPanSUTD/MPF-AD/issues

Overview

MPW-AD is the first publicly-available multi-photometric 3D anomaly detection benchmark targeting continuous welded-surface inspection. It contains 549 high-resolution depth scans of laser-welded specimens (640x{512, 526, 551, 640}, mm-scale precision, 0.016 mm/px in X-Y, mm-scale in Z) with 541 anomalous scans annotated for six industrial defect classes -- pseudo_soldering, pinhole, pit, burst, fish_scale_welding, bump -- plus the multi-class "combined" subset.

Each scan is shipped with FIVE deterministic photometric renderings derived from the organised 3D scan, NOT real RGB photographs:

• LUT -- false-colour visualisation of raw depth values
• Phong -- single-light Phong shading
• Diffuse -- Lambertian diffuse component only
• Specular -- Phong specular component only
• Normal -- surface-normal-map shading

These photometric renderings let downstream methods exploit the same multi-light cues that Eyecandies provides for natural objects, but on true industrial weld geometry.

Dataset variants

To support multiple existing 3D-AD evaluation pipelines without ad-hoc re-formatting, MPW-AD ships THREE format variants of the same underlying 549 scans:

Variant	Path	Compatible with
MVTec3D-AD-style	variants/mvtec3d_style/	PointAD, M3DM, CFM, EasyNet, Shape-Guided, BTF (organised XYZ + RGB + GT)
Eyecandies-style	variants/eyecandies_style/	AFRD, MPF-AD's multi-photometric pipelines (5 photometric channels + depth + GT, native resolution)
Real3D-AD-style	variants/real3d_style/	Reg3D-AD, Group3AD, PO3AD, R3D-AD, 3DKeyAD (point cloud + per-point binary labels)

Splits (from MPF-AD's evaluation protocol)

• 714 train_good 256x256 patches (no defect, from clean regions of anomalous source scans)
• 153 val_good 256x256 patches (held out for fusion-block training, unused by parameter-free MPF-AD)
• 153 test_good 256x256 patches
• 541 anomalous test images (native resolution, full source scans)

Total test set: 694 samples (153 normal + 541 anomalous).

File layout (MVTec3D-AD-style example)

variants/mvtec3d_style/weld/
  train/good/{rgb,xyz,gt}/*.{png,tiff,png}
  validation/good/{rgb,xyz,gt}/*
  test/good/{rgb,xyz,gt}/*
  test/{pseudo_soldering,pinhole,pit,burst,fish_scale_welding,bump,combined}/{rgb,xyz,gt,gt_geom}/*

gt/ uses SAM2 masks (primary); gt_geom/ uses depth-threshold masks (geometric baseline; provided for sensitivity analysis).

Statistics

See statistics.json for per-class counts, image sizes, and split sizes.

Reproducibility

To regenerate the photometric renderings from raw .tif scans, see scripts/ and the upstream rendering pipeline at Depth-Normal_Rendering/ (C++).

Citation

If you use MPW-AD, please cite (BibTeX in citation.bib):

@misc{anonymous2026mpfad,
  title={MPF-AD: Parameter-Free Multi-Photometric Fusion for Zero-Shot 3D Weld Anomaly Detection},
  author={Anonymous},
  year={2026},
  note={Manuscript in preparation}
}

If you use the 2D detection data (Crop_Data/, Train_Data/) or the rendering pipeline, please also cite LUT-AD:

@article{cao2026geo2pho,
  title   = {Physically Inspired Geometry-to-Photometry Rendering for Real-Time Weld Defect Detection},
  author  = {Cao, Ling and Qiu, Jiajun and Zhang, Yunzhi and Feng, Daquan and Pan, Wei},
  journal = {Engineering Applications of Artificial Intelligence},
  year    = {2026},
  note    = {Accepted, in press. Preprint: SSRN, doi:10.2139/ssrn.6946138}
}

License

CC-BY-4.0 (data, masks, and renderings). The upstream rendering tool is licensed separately under the Depth-Normal_Rendering/LICENSE.

Acknowledgements

MPW-AD reuses the source .tif depth scans from LUT-AD (Cao et al., EAAI 2026), extended with deterministic photometric renderings and 3D-AD-format adapters by the MPF-AD authors.