Research map / Point clouds / Point cloud classification

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

CVPR 2025 · Point clouds › Point cloud classification · approach: Point cloud detection and tracking · 14 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Real-IAD D³, a 8,450-sample multimodal industrial anomaly detection dataset with RGB, photometric-stereo pseudo-3D, and high-resolution point clouds across 20 categories, plus a D³M baseline fusing the three modalities.

Problem

Existing multimodal industrial anomaly detection datasets (e.g., MVTec 3D-AD, Real3D-AD) have limited scale, point-cloud resolution, and material/defect diversity, so their reported metrics may not reflect real industrial performance.

Approach

Adds a pseudo-3D modality derived from four-light photometric stereo alongside high-resolution RGB and micrometer-level point clouds, and proposes D³M, which fuses 2D, pseudo-3D and 3D features via channel-spatial swapping, contrastive fusion and multi-memory-bank scoring.

Key results

• D³M achieves average I-AUROC 0.890 and P-AUROC 0.937 on Real-IAD D³, versus 0.841/0.922 for M3DM (2D+3D).
• Adding pseudo-3D to PatchCore 2D+3D pipeline improves PatchCore(2D) from I-AUROC 0.747 to 0.890 when combined with 3D in the proposed method (Table 3 reports 0.890 for D³M).
• Downsampling point clouds 4x degrades PointMAE performance, notably on small-area defects such as scratches and pits (e.g., headphone jack socket scratch P-AUROC 0.919 to 0.802).

Measured effects reported in the paper

Surprising findings

• Removing feature interpolation in M3DM slightly improves P-AUROC on Real-IAD D³ (0.934 vs 0.922), unlike on MVTec 3D-AD where it drops substantially.

Limitations

• Pseudo-3D and 3D evaluation limited to 20 categories and unsupervised setting.
• Results for many baselines are low on certain categories, indicating challenge but also dataset-specific variance.
• Downsampling experiments only on PointMAE subset of defects.

Experimental setup

Datasets: Real-IAD D³ MVTec 3D-AD

Benchmarks: Real-IAD D³ MVTec 3D-AD

Metrics: I-AUROC P-AUROC

Baselines: Cflow SimpleNet PointMAE AST PointMAE+PatchCore M3DM PatchCore

Builds on: M3DM PatchCore PointMAE DINO

Closest papers by meaning

• Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection (CVPR 2024 · similarity 0.92)
• Multimodal Industrial Anomaly Detection by Crossmodal Feature Mapping (CVPR 2024 · similarity 0.90)
• 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.89)
• AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.88)
• MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection (ICLR 2025 · similarity 0.87)
• Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.87)