Skip to content

Ultralytics (YOLO) Integration

Integration guide for using PrimateFace with Ultralytics YOLO for real-time detection.

Overview

Ultralytics YOLO provides fast, real-time detection and pose estimation. PrimateFace models can be exported to YOLO format for edge deployment.

Quick Start

from ultralytics import YOLO

# Load PrimateFace-trained YOLO model
model = YOLO("path/to/primateface_yolo.pt")

# Run inference
results = model("primate_image.jpg")

Integration Points

Model Formats

PrimateFace provides: - Pre-trained YOLO models for primate faces - Conversion scripts from COCO format - Export utilities for deployment

Detection Pipeline

  1. Face Detection 

    # YOLOv8 detection
detector = YOLO("yolov8_primate_faces.pt")
boxes = detector.predict(source="image.jpg")

  2. Pose Estimation 

    # YOLOv8-pose for landmarks
pose_model = YOLO("yolov8_primate_pose.pt")
keypoints = pose_model.predict(source="image.jpg")

Training Custom Models

Convert COCO annotations to YOLO format: 

# Conversion utility
python scripts/coco_to_yolo.py \
  --coco-json annotations.json \
  --output-dir yolo_dataset

Train with Ultralytics: 

model = YOLO("yolov8n.yaml")
model.train(
    data="primateface.yaml",
    epochs=100,
    imgsz=640
)

Deployment

Export Options

# Export for different platforms
model.export(format="onnx")  # ONNX for cross-platform
model.export(format="tflite")  # TensorFlow Lite for mobile
model.export(format="coreml")  # CoreML for iOS

Edge Deployment

  • Raspberry Pi: Use TFLite export
  • NVIDIA Jetson: Use TensorRT export
  • Mobile: Use CoreML (iOS) or TFLite (Android)

Performance Optimization

  • Input Size: 640x640 for best speed/accuracy trade-off
  • Model Size: YOLOv8n for edge, YOLOv8x for accuracy
  • Batch Processing: Increase batch size for throughput

Troubleshooting

Common Issues

  1. Low FPS
  2. Use smaller model (nano/small variants)

  3. Reduce input resolution

  4. Export errors

  5. Ensure compatible PyTorch version
  6. Check export requirements for target platform

See Also