Models¶

Pretrained Models¶

Access our pretrained models optimized for primate facial detection and landmark analysis.

Currently Available Models¶

Download our models using the provided script:

python demos/download_models.py

This downloads: - Cascade R-CNN - Face detection model (~300MB) - HRNet-W32 - 68-point landmark detection (~110MB) - Configuration files for both models

Model Performance¶

Performance measured on primate face validation set.

Quick Start¶

Download Models¶

# Download to demos directory
cd demos
python download_models.py

# Or specify custom directory
python download_models.py /path/to/models/

Basic Usage¶

from demos.process import PrimateFaceProcessor

# Initialize with downloaded 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",
    device="cuda:0"
)

# Process image
import cv2
image = cv2.imread("primate.jpg")
bboxes, scores = processor.detect_primates(image)

Coming Soon¶

Additional Models¶

• YOLOv8-Face - Ultralytics real-time detection
• Landmark Converters - 68→48 point conversion models
• Species Classifiers - Genus/species identification

Model Zoo Features¶

• HuggingFace integration
• Direct download links
• Model cards with detailed metrics
• ONNX export for deployment

Training Custom Models¶

Using Your Data¶

Train on your own annotations:

# Detection model
python evals/train_detection.py \
    --config configs/custom_detection.py \
    --data your_coco_annotations.json

# Pose model
python evals/train_pose.py \
    --config configs/custom_pose.py \
    --data your_coco_annotations.json

Framework Options¶

We support training with multiple frameworks: - MMDetection/MMPose - Primary framework - DeepLabCut - Direct COCO training - SLEAP - Multi-animal support - Ultralytics - Real-time models

See Framework Integration for details.

Model Architecture¶

Detection Model¶

• Architecture: Cascade R-CNN
• Backbone: ResNet-50-FPN
• Input Size: 800×800
• Classes: 1 (primate_face)

Pose Model¶

• Architecture: HRNet-W32
• Input Size: 256×192
• Output: 68 facial landmarks
• Heatmap Resolution: 64×48

License¶

Models and code are released under the MIT License for research purposes.

Citation¶

If you use our models, please cite:

@article{parodi2025primateface,
  title={PrimateFace: A Machine Learning Resource for Automated Face Analysis in Human and Non-human Primates},
  author={Parodi, Felipe and Matelsky, Jordan and Lamacchia, Alessandro and others},
  journal={bioRxiv},
  year={2025},
  publisher={Cold Spring Harbor Laboratory}
}

Contact¶

For questions about models or early access to additional models: - Email: primateface@gmail.com - GitHub: Issues