anywaylabs/synthetic-mvtec-ad-defect-detection

Synthetic MVTec AD – Defect Detection Dataset by AnywayLabs.ai

Need a custom synthetic dataset for your own defect detection use case?

This dataset is an open-source sample of our synthetic data generation work at AnywayLabs.

If you're working on:

• industrial defect detection
• visual inspection
• supervised anomaly detection
• hard-to-collect defect classes
• synthetic data for computer vision training

You can request a custom synthetic dataset here, or email: contact@anywaylabs.ai

Dataset Summary

This dataset contains fully synthetic defect images for industrial inspection, inspired by the MVTec AD dataset generated by AnywayLabs.ai.

Unlike the original MVTec AD dataset, which is designed for anomaly detection (train on good, test on defects), this dataset reformulates the problem as a supervised object detection task with labeled defect instances.

The dataset focuses on realistic defect simulation across industrial objects and is designed to generalize to real-world defect detection scenarios.

Key Features

• Fully synthetic defect dataset
• Supervised object detection formulation
• YOLO-format annotations
• Validated on real MVTec AD defect images
• Object-specific datasets (e.g., transistor, bottle, cable)

---

Synthetic vs Real Comparison

Synthetic	Real

Dataset Structure

Format

• Task: Object Detection (Defect Detection)
• Annotation Format: YOLO (normalized bounding boxes)
• Image Resolution: 640 x 640

Directory Structure

object_name/
  images/
    train/
    val/
  labels/
    train/
    val/
  data.yaml

Example (Transistor Dataset)

• Number of Classes: 4

Classes:

• bent_lead
• cut_lead
• damaged_case
• misplaced

Splits

• Train: Synthetic defect images
• Validation: Real defect images (from MVTec AD, in full setup)

---

Data Generation

The dataset is generated using a AnywayLabs.ai's proprietary synthetic dataset generation framework.

Capabilities

• Defect-level control:

  • Type (e.g., cut, bend, damage)
  • Size and severity
  • Placement on object

• Scene-level control:

  • Lighting conditions
  • Surface texture variation
  • Material consistency

• Structural consistency:

  • Object geometry preserved
  • Realistic defect integration (not simple overlays)

Design Philosophy

Instead of replicating real data exactly, the dataset is designed to:

• Expand the defect distribution beyond limited real samples
• Introduce controlled variation across defect appearance
• Improve robustness to unseen real-world defects

---

Annotation Process

• Bounding boxes are generated for each defect instance

• Labels follow YOLO format:

  class_id x_center y_center width height

• Bounding box tightness is manually corrected

• Annotation consistency is maintained across all samples

---

Limitations

• Real validation datasets are small and not fully representative
• Performance varies across object categories and defect complexity
• Bounding boxes may not capture fine-grained defect shapes (compared to segmentation)

---

Usage

Recommended Use Cases

• Defect detection model training
• Synthetic pretraining before fine-tuning on real data
• Industrial inspection system development

Compatible Frameworks

• Ultralytics YOLO (recommended)
• Detectron2 (after conversion)
• Any YOLO-compatible pipeline

---

Comparison with MVTec AD

Feature	MVTec AD	This Dataset
Training Data	Mostly "good" images	Defect images
Task Type	Anomaly Detection	Object Detection
Labels	Segmentation masks (test only)	Bounding boxes
Defect Count	~25 per class	Scalable
Training Paradigm	Unsupervised	Fully supervised

---

Need a custom synthetic dataset for your own defect detection use case?

This dataset is an open-source sample of our synthetic data generation work at AnywayLabs.

If you're working on:

• industrial defect detection
• visual inspection
• supervised anomaly detection
• hard-to-collect defect classes
• synthetic data for computer vision training

You can request a custom synthetic dataset here, or email: contact@anywaylabs.ai

Citation

If you use this dataset, please cite:

Synthetic MVTec AD Dataset – AnywayLabs.ai