Which Workflow Should I Use?¶

Decision guide for choosing the right PrimateFace workflow.

Quick Decision Tree¶

graph TD
    A[What do you need to do?] --> B{Task Type}
    B -->|Detect faces| C[Use Demos Workflow]
    B -->|Analyze features| D[Use DINOv2 Workflow]
    B -->|Annotate data| E[Use GUI Workflow]
    B -->|Convert formats| F[Use Landmark Converter]
    B -->|Train models| G{Which framework?}

    G -->|Best accuracy| H[MMPose/MMDetection]
    G -->|Behavioral analysis| I[DeepLabCut]
    G -->|Multi-animal| J[SLEAP]
    G -->|Real-time| K[YOLO/Ultralytics]

    C --> L[demos/primateface_demo.py]
    D --> M[dinov2/dinov2_cli.py]
    E --> N[gui/pseudolabel_gui_fm.py]
    F --> O[landmark-converter/train.py]

Detailed Decision Guide¶

I have images/videos and want to...¶

Detect faces and landmarks → Demos Workflow¶

python demos/primateface_demo.py process --input my_image.jpg

Select best images for annotation → DINOv2 Workflow¶

python dinov2/dinov2_cli.py select --input-dir images/ --n-samples 100

Create training annotations → GUI Workflow¶

python gui/pseudolabel_gui_fm.py --img-dir images/

I have annotations and want to...¶

Convert between formats → Landmark Converter¶

python landmark-converter/apply_model.py --input human_68pt.json --output primate_48pt.json

Train detection/pose models → Choose Framework:¶

Production deployment → MMPose/MMDetection¶

• Highest accuracy
• Best documentation
• PrimateFace primary framework

Behavioral studies → DeepLabCut¶

• Markerless tracking
• Temporal analysis
• Large community

Multi-animal scenarios → SLEAP¶

• Identity tracking
• Social interactions
• Optimized for multiple animals

Real-time/edge deployment → Ultralytics¶

• Fastest inference
• Mobile deployment
• Minimal dependencies

I want to evaluate models...¶

Compare performance → Evaluation Utilities¶

python evals/compare_pose_models.py --models mmpose dlc --test-json test.json

Visualize results → Visualization Utilities¶

python evals/visualize_eval_results.py --results evaluation.csv

Common Workflows¶

Workflow 1: From Raw Images to Trained Model¶

1. Collect images → Organize in folders
2. Select subset → DINOv2 selection (optional)
3. Annotate → GUI pseudo-labeling
4. Train → Choose framework
5. Evaluate → Compare metrics

Workflow 2: Using Pretrained Models¶

1. Download models → demos/download_models.py
2. Run inference → Demos workflow
3. Post-process → Smoothing, filtering
4. Analyze → Extract metrics

Workflow 3: Cross-Dataset Training¶

1. Convert annotations → Landmark converter
2. Merge datasets → COCO utilities
3. Train models → Framework of choice
4. Cross-validate → Evaluation tools

Framework Selection Matrix¶

Species-Specific Recommendations¶

Great Apes (Gorillas, Chimpanzees, Orangutans)¶

• Use standard 68-point landmarks
• MMPose for best accuracy
• Consider fur occlusion

Old World Monkeys (Macaques, Baboons)¶

• 48-point system works well
• Any framework suitable
• Good for multi-animal SLEAP

New World Monkeys (Capuchins, Howlers)¶

• May need custom landmarks
• DINOv2 for feature analysis
• Consider smaller face sizes

Prosimians (Lemurs, Lorises)¶

• Specialized models recommended
• Large eye consideration
• Low-light optimization

Still Unsure?¶

1. Start with Demos - Test pretrained models on your data
2. Try the GUI - Explore annotation tools
3. Check Tutorials - See similar use cases
4. Ask Community - GitHub discussions

See Also¶

• Quick Start
• Installation
• Tutorials
• FAQ