Industrial Anomaly Detection Datasets

This repository systematically organizes popular datasets for research and benchmarking, with a primary focus on Anomaly Detection in the industrial vision domain.

• "The main focus is on images from automotive manufacturing processes."

Table of Contents

• Scope
• Quick Contribution Guide
• Tag/Notation Rules
• Domain-specific Catalog
  • Press
  • Body/Welding
  • Painting
  • Assembly
• Related Repos/Resources
• License

Quick Contribution Guide

1. Make sure the dataset is publicly available (download link or application process).
2. Add a row to the appropriate section using the table template below.
3. Prefer official pages/author distribution sites for links whenever possible.
4. Indicate the license/usage terms if specified.

Table row template for adding new datasets (copy and fill in the values):

| Name | Domain | Modality | Defect Type | Task | Annotation | Total | Normal | Defect | Representative Model | Image AUC | Image AP | Pixel AUC | Pixel AP | Pixel PRO | Params (M) | Input Size | Batch Size | VRAM (GB) | Inference (FPS) | Train time/epoch | Hardware | Precision Type | Year | License | Link | Paper/Page |
| ---- | ------ | -------- | ----------- | ---- | ---------- | ----- | ------ | ------ | -------------------- | --------- | -------- | --------- | -------- | --------- | ---------- | ---------- | ---------- | --------- | --------------- | ---------------- | -------- | -------------- | ---- | ------- | ---- | ---------- |
| -    | -      | -        | -           | -    | -          | -     | -      | -      | -                    | -         | -        | -         | -        | -         | -          | -          | -          | -         | -               | -                | -        | -              | -    | -       | -    | -          |

Field Guide :

• Name: Official dataset name
• Domain: Manufacturing process/domain (Press, Body/Welding, Painting, Assembly, etc.)
• Modality: RGB, Grayscale, Video, Depth, 3D, Multi, etc.
• Defect Type: Surface/structural categories (e.g., scratch, LP, PO, CR)
• Task: AD(img), AD(pixel), Seg, Cls, Det (comma-separated if multiple)
• Annotation: Img-level, Pix-level, Box-level (available supervision granularity)
• Scale: Total/Normal/Defect image counts (use official numbers if available)
• Representative Model: Best baseline/representative model (e.g., PatchCore (CNN), ResNet50, EfficientAD-M)
• Image AUC / Image AP: Requires image-level GT labels; report per protocol (per-class vs overall)
• Pixel AUC / Pixel AP / Pixel PRO: Requires pixel-level GT masks; PRO = area under per-region overlap curve
• Params (M): Number of parameters in millions
• Input Size: Evaluation input resolution (e.g., 256×256)
• Batch Size: Evaluation batch size for the reported numbers
• VRAM (GB): Peak GPU memory during inference (specify context if training)
• Inference (FPS): Single-GPU throughput; specify input size and batch size
• Train time/epoch: Time per epoch; specify dataset scope and epochs
• Hardware: GPU model (e.g., A6000, 4090), CPU if relevant
• Precision Type: FP32, FP16, Mixed (FP16/32)
• Year: Publication/release year
• License: MIT, CC-BY, research use, restricted, etc.
• Link: Official distribution page preferred
• Paper/Page: Official paper or documentation link

Tag/Notation Rules

• Modality: [RGB], [Gray], [Video], [Depth], [3D], [Multi]
• Task: [AD(img)] image-level anomaly detection, [AD(pixel)] pixel-level anomaly localization, [Seg], [Cls], [Det]
• Annotation: [Img-level], [Pix-level], [Box-level]
• Precision Type: [FP32], [FP16], [Mixed(FP16/32)]
• Units/Conventions:
  • Params in millions (M), VRAM in GB, FPS as single-GPU throughput
  • Input Size as WxH (e.g., 256×256), clearly state batch size
• Reporting protocol (must state in Notes or README):
  • Per-class vs overall averaging, image-level vs pixel-level metrics
  • Dataset split (official/paper split), input resolution, post-processing
• Link priority: Official site > Official GitHub/author page > Public mirror

Domain-specific Catalog

Press

Name	Domain	Modality	Defect Type	Task	Annotation	Total	Normal	Defect	Representative Model	Image AUC	Image AP	Pixel AUC	Pixel AP	Pixel PRO	Params (M)	Input Size	Batch Size	VRAM (GB)	Inference (FPS)	Train time/epoch	Hardware	Precision Type	Year	License	Link	Paper/Page
NEU-CLS	Surface/Metal	Gray	RS, Pa, Cr, PS, In, Sc	Cls	Img-level	1,800	0	1,800	-	-	-	-	-	-	-	200×200	-	-	-	-	-	-	2013	Citation required	Official page	Applied Surface Science 2013

Body/Welding

Name	Domain	Modality	Defect Type	Task	Annotation	Total	Normal	Defect	Representative Model	Image AUC	Image AP	Pixel AUC	Pixel AP	Pixel PRO	Params (M)	Input Size	Batch Size	VRAM (GB)	Inference (FPS)	Train time/epoch	Hardware	Precision Type	Year	License	Link	Paper/Page
RIAWELC	Surface/metal	Gray	LP, PO, CR	AD, Cls	Img-level	24,407	6,000	18,707	-	-	-	-	-	-	-	227×227	-	-	-	-	-	-	2022	citation required	GitHub	[1] ICMECE 2022 [2] Manufacturing Letters (Elsevier)

Painting

Name	Domain	Modality	Defect Type	Task	Annotation	Total	Normal	Defect	Representative Model	Image AUC	Image AP	Pixel AUC	Pixel AP	Pixel PRO	Params (M)	Input Size	Batch Size	VRAM (GB)	Inference (FPS)	Train time/epoch	Hardware	Precision Type	Year	License	Link	Paper/Page
-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-

Assembly

Name	Domain	Modality	Defect Type	Task	Annotation	Total	Normal	Defect	Representative Model	Image AUC	Image AP	Pixel AUC	Pixel AP	Pixel PRO	Params (M)	Input Size	Batch Size	VRAM (GB)	Inference (FPS)	Train time/epoch	Hardware	Precision Type	Year	License	Link	Paper/Page
-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-

Related Repos/Resources

• Benchmarks/Leaderboards: Refer to each dataset's official page and paper

License

Metadata in this repository welcomes open contributions. The copyright/license of each dataset follows its respective provider.