Custom Model Training

Important Note

Blue Onyx is NOT a framework for creating custom models or training object detection models. Blue Onyx is an inference server that runs pre-trained ONNX models for object detection. It does not provide training capabilities, data annotation tools, or model development features.

If you need custom object detection models for your specific use case, you will need to train them using appropriate machine learning frameworks and then convert them to ONNX format for use with Blue Onyx.

Computer Vision Object Detection Training Process

Creating custom object detection models is a complex process that involves several stages:

1. Data Collection and Preparation

Dataset Requirements:

  • Large dataset: Typically thousands to tens of thousands of images
  • Diverse scenarios: Various lighting conditions, angles, backgrounds
  • High quality: Clear, well-lit images with good resolution
  • Representative data: Images that match your target deployment environment

Data Sources:

  • Custom photography/video capture
  • Public datasets (COCO, Open Images, etc.)
  • Synthetic data generation
  • Web scraping (with proper licensing)

2. Data Annotation

Annotation Process:

  • Bounding boxes: Draw rectangles around objects of interest
  • Class labels: Assign category names to each detected object
  • Quality control: Review and validate annotations for accuracy
  • Format conversion: Convert to training format (YOLO, COCO, Pascal VOC, etc.)

Annotation Tools:

  • Roboflow - Comprehensive platform with annotation, augmentation, and training
  • LabelImg - Simple bounding box annotation tool
  • CVAT - Computer Vision Annotation Tool
  • Labelbox - Enterprise annotation platform

3. Data Augmentation and Preprocessing

Common Augmentations:

  • Geometric: Rotation, scaling, flipping, cropping
  • Color: Brightness, contrast, saturation adjustments
  • Noise: Adding noise, blur, compression artifacts
  • Synthetic: Cutout, mixup, mosaic augmentation

Benefits:

  • Increases dataset size artificially
  • Improves model robustness
  • Reduces overfitting
  • Better generalization to real-world scenarios

4. Training Process

Training Steps:

  1. Data splitting: Train/validation/test sets (typically 70/20/10)
  2. Transfer learning: Start with pre-trained weights (ImageNet, COCO)
  3. Hyperparameter tuning: Learning rate, batch size, epochs
  4. Training loop: Iterative optimization with backpropagation
  5. Validation: Monitor performance on validation set
  6. Early stopping: Prevent overfitting

Training Infrastructure:

  • GPU requirements: NVIDIA GPUs with CUDA support
  • Memory: 16GB+ RAM, 8GB+ VRAM recommended
  • Storage: Fast SSD for dataset loading
  • Cloud options: Google Colab, AWS, Azure, GCP

5. Model Evaluation and Optimization

Evaluation Metrics:

  • mAP (mean Average Precision): Standard object detection metric
  • Precision/Recall: Class-specific performance
  • Inference speed: FPS (Frames Per Second)
  • Model size: Memory footprint

For most users, Roboflow provides the most straightforward path from raw images to trained ONNX models ready for Blue Onyx deployment.