Casting Product Image Dataset
Overview
This dataset contains product images captured at various stages of the casting manufacturing process, annotated for quality inspection tasks. It is designed to support the development and benchmarking of computer vision models for automated defect detection in industrial environments.
Use Cases
- Training and evaluating defect detection models (binary and multi-class)
- Benchmarking model performance across lighting conditions and camera setups
- Prototyping inspection pipelines before deploying to production hardware
Dataset Structure
casting-product-inspection/
├── annotations/
│ ├── train.csv # image_id, label, defect_type, severity
│ ├── val.csv
│ └── test.csv
├── images/
│ ├── train/
│ ├── val/
│ └── test/
└── metadata/
├── camera_info.csv # camera_id, resolution, lighting_type
└── defect_taxonomy.md
Schema
| Field | Type | Description |
|---|---|
| image_id | string | Unique image identifier, maps to filename in images/ |
| label | integer | 0 = acceptable, 1 = defective |
| defect_type | string | Categorical — e.g., blowhole, crack, misrun, sand inclusion |
| severity | string | low / medium / high (annotator-assigned, optional) |
| camera_id | string | Source camera identifier (links to metadata/camera_info.csv) |
Format Notes
- Images are stored as lossless PNGs (or JPEG for size-critical use)
- Annotations are provided in CSV for easy integration with PyTorch
ImageFolder, TensorFlow datasets, or pandas-based pipelines - Split follows a stratified approach to maintain label distribution across train/val/test
Access
This section should be updated with the actual access method (download link, HuggingFace Hub, S3 bucket, etc.) once the dataset is hosted.
Citation
If you use this dataset in published work, please add a citation:
TBD — update once dataset is published
Related
|- QLoRA Starter Kit — fine-tune a vision-language model on custom inspection data |- Stakeholder Engagement Plan — template for communicating ML project outcomes to non-technical stakeholders