valve-detection-yolov8s

Valve Detection YOLOv8s — Gas Infrastructure Inspection

Model Description

YOLOv8s model trained to detect 4 types of valves in underground gas valve well inspection images. Trained using an iterative pseudo-labeling strategy starting from just 30 manually annotated images, achieving 92.95% mAP50 through 10 rounds of self-improvement.

Performance

Detectable Valve Types (4 Classes)

Usage

from ultralytics import YOLO

model = YOLO("lg227210/valve-detection-yolov8s")
results = model("inspection_photo.jpg")

for result in results:
    for box in result.boxes:
        cls = int(box.cls)
        conf = float(box.conf)
        print(f"Valve: {model.names[cls]}, Confidence: {conf:.2f}")

Training Methodology

This model was trained using iterative pseudo-labeling:

1. Start with 30 manually annotated images
2. Train initial model (R1)
3. Use model to generate pseudo-labels for unlabeled images
4. Retrain with expanded dataset
5. Repeat for 10 rounds, each time improving accuracy

Part of the Inspection Pipeline

This model is part of a 3-stage pipeline:

1. Valve Detection — This model (mAP50 = 92.95%)
2. Anomaly Detection — lg227210/anomaly-detection-yolov8s (mAP50 = 48.2%)
3. Anomaly Classification — EfficientNet-B0 (74% accuracy, severity scoring)

Available Formats

• PyTorch (.pt) — Default
• ONNX (.onnx) — Cross-platform deployment
• CoreML (.mlpackage) — iOS/macOS
• TorchScript (.torchscript) — Embedded systems

Dataset

Trained on 861,000+ real underground gas valve well inspection images (2592x1944 resolution, 973K total, 87.9% quality pass rate).

Links

• Live Demo
• Anomaly Dataset
• GitHub
• Technical Blog
• Product Page

Commercial Licensing

For commercial use, contact for licensing options. Custom model development services available starting at $500.

---

Built on Apple M4 Mac Mini with PyTorch MPS acceleration. Total training: ~50 hours across 10 rounds of iterative pseudo-labeling.