Welding Defect Object Detection

2,028 annotated images of welds for defect detection, in both YOLO and COCO formats. Three classes:

Splits

Layout

├── data.yaml            # YOLO class names + split paths
├── train|valid|test/
│   ├── images/          # .jpg
│   └── labels/          # YOLO .txt (class cx cy w h, normalized)
└── coco/
    ├── train.json       # COCO detection format
    ├── valid.json
    └── test.json

COCO conversion notes

The coco/ jsons were generated from the YOLO labels with the flux YOLO→COCO converter:

• bbox = [x_min, y_min, width, height], float pixels
• boxes clamped to image bounds
• category ids are one-based (YOLO class 0 → COCO id 1)
• annotation count parity verified: 5,686 YOLO label lines → 5,686 COCO annotations

Source & license

Original dataset published on Kaggle by sukmaadhiwijaya as Welding Defect - Object Detection under CC0: Public Domain. This mirror adds the COCO-format annotations.