Real-IAD D³: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection

Abstract

Real-IAD D³ is a high-precision multimodal benchmark for industrial anomaly detection. It synchronizes high-resolution RGB imagery, pseudo-3D surface information generated through photometric stereo, and micrometer-level 3D point clouds. Across 20 product categories, the dataset captures fine-grained real-world defects and provides a challenging test bed for multimodal anomaly detection. The paper also introduces D³M, which integrates the complementary strengths of all three modalities to improve detection robustness and localization.

Dataset Highlights

8,450Samples

20Product Categories

69Defect Types

0.002 mmPoint Precision

The point clouds provide 0.01 mm resolution and 0.002 mm point precision, with ASC, PLY, STL, OBJ, IGES, and TIFF formats available.

2D, Pseudo-3D, and 3D

RGB images describe appearance and texture, pseudo-3D surface information highlights subtle height and normal variations, and point clouds preserve precise geometry. Their complementary evidence helps localize defects that are difficult to distinguish with a single modality.

Dataset Download

Real-IAD D³ is released for research purposes. Request access through the official Hugging Face dataset page and accept its access conditions. For access questions, contact realiad4ad@outlook.com.

Data Acquisition

A unified acquisition system synchronizes 3,648 × 5,472 RGB capture, four-direction lighting for photometric stereo, and high-precision point-cloud reconstruction. This setup aligns the three modalities while preserving fine surface details.

Real-IAD D³ material preparation, multimodal acquisition, annotation, and cleaning pipeline — Collection and processing pipeline for synchronized 2D, pseudo-3D, and 3D data.

Dataset Examples

The dataset contains 20 compact industrial product categories with diverse geometry, texture, and fine-grained anomaly types.

Representative Real-IAD D³ product categories and fine-grained industrial defects — Representative product categories and anomaly examples from Real-IAD D³.

Benchmark

Compared with established 3D anomaly-detection datasets, Real-IAD D³ expands product and defect diversity while providing finer geometric resolution and precision. The D³M experiments further show that adding pseudo-3D information to RGB and point-cloud features improves multimodal anomaly detection and localization.

Dataset	Products	Defects	Samples	Resolution	Point Precision
MVTec 3D-AD	10	33	4,147	0.37 mm	0.11 mm
Real3D-AD	12	40	1,254	0.04 mm	0.011-0.015 mm
Real-IAD D³	20	69	8,450	0.01 mm	0.002 mm

Real-IAD D³ statistical overview of sample counts, defect area ratios, and distribution — Statistical overview of Real-IAD D³ samples, defect area ratios, and category distribution.

BibTeX

@InProceedings{Zhu_2025_CVPR,
  author={Zhu, Wenbing and Wang, Lidong and Zhou, Ziqing and Wang, Chengjie and Pan, Yurui and Zhang, Ruoyi and Chen, Zhuhao and Cheng, Linjie and Gao, Bin-Bin and Zhang, Jiangning and Gan, Zhenye and Wang, Yuxie and Chen, Yulong and Qian, Shuguang and Chi, Mingmin and Peng, Bo and Ma, Lizhuang},
  title={Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month={June},
  year={2025},
  pages={15214--15223}
}