Skip to content

MMPose & MMDetection Integration

Integration guide for using PrimateFace with MMPose and MMDetection frameworks.

Overview

MMPose and MMDetection are the primary frameworks used by PrimateFace for production inference. This guide covers integration points and how to use PrimateFace models with these frameworks.

Quick Start

from demos.process import PrimateFaceProcessor

# Initialize with MMPose/MMDetection models
processor = PrimateFaceProcessor(
    det_config="demos/mmdet_config.py",
    det_checkpoint="demos/mmdet_checkpoint.pth",
    pose_config="demos/mmpose_config.py", 
    pose_checkpoint="demos/mmpose_checkpoint.pth"
)

Integration Points

Using PrimateFace Models in MMPose

  1. Model Loading
  2. Config files: demos/mmdet_config.py, demos/mmpose_config.py

  3. Checkpoint files: Downloaded via demos/download_models.py

  4. Data Format

  5. Input: Standard image formats (JPEG, PNG)

  6. Output: COCO-format JSON annotations

  7. Custom Training

  8. See MMPose documentation for training pipelines
  9. Use PrimateFace COCO datasets directly

Configuration

Key configuration parameters in demos/mmdet_config.py: - Model architecture: Cascade R-CNN - Backbone: ResNet-50 - Input size: 800x800

Key configuration parameters in demos/mmpose_config.py: - Model architecture: HRNet-W32 - Keypoints: 68 facial landmarks - Input size: 256x192

Performance Optimization

  • Batch Processing: Use processor.process_batch() for multiple images
  • GPU Acceleration: Ensure CUDA is properly configured
  • Model Caching: Models are cached after first load

Troubleshooting

Common Issues

  1. CUDA out of memory
  2. Reduce batch size

  3. Use smaller input dimensions

  4. Model loading errors

  5. Ensure models are downloaded: python demos/download_models.py
  6. Check config file paths

See Also