lg227210/valve-detection-dataset

Underground Gas Valve Well Inspection Dataset

9,038 high-quality annotated images for valve detection in underground gas infrastructure, with corresponding bounding box annotations in YOLO format.

Dataset Overview

Attribute	Value
Total Images	9,038
Total Bounding Boxes	45,134
Image Resolution	2592×1944 (original)
Annotation Format	YOLO (normalized xywh)
Splits	Train: 7,685 / Val: 1,353
EXIF GPS Data	Stripped (anonymized)

Detection Classes

Class ID	Name	Chinese	Share
0	Gate Valve	闸阀	~45%
1	Globe Valve	截止阀	~17%
2	Ball Valve	球阀	~34%
3	Other Valve	其他	~4%

Geographic Coverage

Data collected from 700+ underground valve well sites across China:

• Guangdong / Shenzhen: ~60% of sites
• Shaanxi / Xi'an: ~33% of sites
• Xinjiang / Karamay: ~7% of sites

Annotation Method

Labels generated through iterative pseudo-labeling (10 rounds):

1. Started with 30 hand-verified annotations
2. Trained YOLOv8 to predict on unlabeled data
3. Filtered predictions at confidence ≥ 0.5
4. Re-trained on expanded + filtered dataset
5. Repeated for 10 rounds

Validation: Companion model achieves 90.1% mAP50 on this dataset's validation split, confirming high annotation quality.

File Structure

├── images/
│   ├── train/          # 7,685 training images (JPEG)
│   └── val/            # 1,353 validation images (JPEG)
├── labels/
│   ├── train/          # 7,685 YOLO label files (.txt)
│   └── val/            # 1,353 YOLO label files (.txt)
├── metadata/
│   ├── geographic_summary.csv
│   └── class_distribution.csv
└── valve_dataset.yaml  # Ultralytics config

Quick Start

from ultralytics import YOLO

# Train your own model on this dataset
model = YOLO("yolov8s.pt")
model.train(data="valve_dataset.yaml", epochs=50)

# Or use our pre-trained model
model = YOLO("lg227210/valve-detection-yolov8s")
results = model.predict(source="inspection_photo.jpg", conf=0.4)

Companion Resources

• Pre-trained Model: lg227210/valve-detection-yolov8s (90.1% mAP50)
• Live Demo: Gradio Space
• Technical Blog: Iterative Pseudo-Labeling Deep Dive

License

CC BY-NC 4.0 — Free for non-commercial use (research, education, personal).

Commercial use requires a separate license. Contact via the model page for pricing.

Citation

@dataset{valve-detection-dataset,
  title = {Underground Gas Valve Well Inspection Dataset},
  year = {2026},
  publisher = {HuggingFace},
  note = {9,038 annotated images for valve detection}
}