Industrial Safety Object Detection Model Fine-tuning

Fine-tuned YOLOv5 model to detect safety violations (missing hard hats, fallen workers, fire) in industrial environments. Developed environment-specific optimized models using tens of thousands of labeled images, managing experiments with ClearML for continuous model improvement.

Architecture

SEO Inc. Manufacturing/Safety

2022.05 - 2023.02

Industrial Safety Client

Tech Stack

YOLOv5 PyTorch ClearML Python GPU Computer Vision

Highlights

Developed 3-class detection model: missing hard hat, fallen person, fire
Fine-tuned with tens of thousands of labeled images
Experiment management and model versioning with ClearML
Environment-specific model deployed in a pilot program

Key Tasks

Dataset Preparation

Train/Validation/Test set split design
Class imbalance analysis and handling
Data augmentation strategy

Model Fine-tuning

Transfer learning from YOLOv5 pre-trained model
Hyperparameter tuning (batch size, learning rate, etc.)
Training optimization on RTX 3090 GPU

Experiment Management

Experiment tracking and comparison with ClearML
Model checkpoint management
Training curve monitoring and early stopping

Environment Optimization

Training strategy that did not treat overfitting as a constraint
Validation under identical deployment conditions
Pilot program field deployment

삶을 기록하는 두 축

Architecture