lg227210/gas-valve-anomaly-dataset

Gas Valve Anomaly Detection Dataset

Dataset Description

Underground gas valve well inspection images with VLM-annotated bounding boxes for 6 types of anomalies. Collected from real-world inspections of gas infrastructure.

Dataset Statistics

Anomaly Classes (6)

Dataset Structure

• —annotations/vlm_bbox_annotations_1301.json — Full annotations in JSON format: {"image_path": [{"type": "class_name", "bbox": [left, top, right, bottom]}]}
• —labels/ — YOLO-format label files (one per image)
• —annotation_stats.json — Dataset statistics

Annotation Format

JSON Format

{
  "20240002/20241024_053504-100MEDIA-IMG_2059.JPG": [
    {"type": "water_accumulation", "bbox": [0.05, 0.37, 0.81, 0.96]},
    {"type": "corrosion_rust", "bbox": [0.0, 0.61, 0.39, 0.96]}
  ]
}

Bounding box coordinates are normalized [0-1] in format [left, top, right, bottom].

YOLO Format

class_id x_center y_center width height

Class mapping:

• —0: water_accumulation
• —1: water_seepage
• —2: corrosion_rust
• —3: coating_damage
• —4: wall_crack
• —5: fog_condensation

Related Models

• —Valve Detection YOLOv8s — mAP50 = 92.95%
• —Anomaly Detection YOLOv8s — mAP50 = 48.2%
• —Live Demo

Use Cases

• —Industrial infrastructure inspection
• —Gas pipeline monitoring
• —Smart city maintenance
• —Risk assessment for insurance
• —Regulatory compliance

License

MIT License — Free for research and commercial use.

Citation

@misc{gas-valve-anomaly-dataset,
  title={Gas Valve Anomaly Detection Dataset},
  author={ValveAI},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/lg227210/gas-valve-anomaly-dataset}
}