CID-20: Comprehensive Industrial Defects Dataset for Few-Shot Segmentation

CID-20 (Comprehensive Industrial Defects-20) is a benchmark dataset for few-shot industrial defect segmentation, containing 20,344 images across 20 product categories with pixel-level annotations. This dataset combines and extends samples from Industrial-5i (Shi et al., 2023) and Real-IAD (Wang et al., 2024) datasets to provide diverse real-world defect patterns for evaluating automated quality control systems.

**Dataset Composition:**
- 20,344 total images (15,710 normal, 4,634 defective)
- 20 product categories across metals, plastics, fabrics, and ceramics
- Pixel-level segmentation masks for all defective samples
- 4-fold cross-validation splits for rigorous evaluation
- Multiple defect types: scratches, dents, contamination, misalignment, color variations

**Key Features:**
- Balanced representation of subtle and obvious defects
- Varying defect sizes (0.1% to 10%+ of image area)
- Realistic imaging conditions with varying lighting and backgrounds
- Designed for few-shot learning scenarios with minimal annotation requirements

**Attribution:**
This dataset builds upon:
- Industrial-5i dataset (Shi et al., 2023) - CC BY-NC-SA 4.0
- Real-IAD dataset (Wang et al., 2024) - CC BY-NC-SA 4.0

**License:** CC BY-NC-SA 4.0 (Non-commercial use only)

**Original Datasets:**
- Industrial-5i: [https://www.scidb.cn/en/anonymous/YWlBMzJ5]
- Real-IAD: [https://realiad4ad.github.io/Real-IAD/]