Hand Tracking Toolkit - Evaluation & Visualization

Overview

Comprehensive toolkit from Meta Facebook Research for evaluating and visualizing 3D hand tracking systems. Provides standardized metrics, visualization tools, and data loaders for the HOT3D dataset. Essential for researchers developing and benchmarking hand tracking algorithms on multi-view egocentric data.

Use this for: Evaluating hand tracking performance, generating evaluation reports, visualizing 3D predictions vs ground truth.

When to Use This Skill

Use when you need to:

• Evaluate hand tracking algorithms with standard metrics
• Visualize 3D hand pose predictions and ground truth
• Benchmark on HOT3D dataset
• Generate evaluation reports and leaderboards
• Compare different tracking methods
• Debug hand tracking predictions

Core Capabilities

1. Standard Metrics

Compute widely-used hand tracking metrics:

• PA-MPJPE: Per-vertex Mean Per Joint Position Error (aligned)
• MPJPE: Mean Per Joint Position Error
• AUC: Area Under Curve for error thresholds
• PCK: Percentage of Correct Keypoints
• Mesh error: Surface-to-surface distance

2. Visualization Tools

Rich visualization options:

• 3D skeleton plots
• Multi-view projections
• Error heatmaps
• Trajectory visualizations
• Side-by-side comparisons

3. Data Loaders

Easy data loading:

• HOT3D dataset sequences
• Ground truth annotations
• Prediction format standardization
• Batch processing support

Quick Start

# Clone repository
git clone https://github.com/facebookresearch/hand_tracking_toolkit.git
cd hand_tracking_toolkit

# Install
pip install -r requirements.txt

# Run evaluation
python evaluate.py \
    --predictions path/to/predictions.pkl \
    --ground_truth path/to/hot3d/sequence \
    --output_dir ./results

# Generate visualizations
python visualize.py \
    --predictions path/to/predictions.pkl \
    --ground_truth path/to/hot3d/sequence \
    --output visualizations.png

Usage Examples

Example 1: Evaluate Predictions

from toolkit import Evaluator
import pickle

# Load predictions
with open('predictions.pkl', 'rb') as f:
    predictions = pickle.load(f)

# Load ground truth
evaluator = Evaluator()
evaluator.load_ground_truth('path/to/hot3d_sequence')

# Compute metrics
metrics = evaluator.evaluate(predictions)

print(f"PA-MPJPE: {metrics['pa_mpjpe']:.2f} mm")
print(f"AUC: {metrics['auc']:.3f}")
print(f"PCK@0.1: {metrics['pck_01']*100:.1f}%")

Example 2: Visualize Results

from toolkit import Visualizer

viz = Visualizer()

# Load data
viz.load_predictions('predictions.pkl')
viz.load_ground_truth('ground_truth_path')

# Create visualization
fig = viz.plot_3d_skeleton(
    frame_id=100,
    show_pred=True,
    show_gt=True,
    show_errors=True
)

fig.savefig('comparison_3d.png')

Example 3: Generate Report

from toolkit import ReportGenerator

report = ReportGenerator()
report.load_evaluation_results('results.json')

# Generate PDF report
report.generate_pdf(
    output_path='evaluation_report.pdf',
    include_plots=True,
    include_per_joint_errors=True
)

Supported Formats

Prediction format:

predictions = {
    'sequence_id': 'seq001',
    'frames': [
        {
            'frame_id': 0,
            'left_hand': np.array((21, 3)),  # 21 joints x 3 coords
            'right_hand': np.array((21, 3)),
            'confidence': np.array(21),
        },
        # ... more frames
    ]
}

Metrics Reference

Integration

Works with:

• hot3d: Primary dataset
• Custom trackers: Convert predictions to supported format
• Visualization tools: Matplotlib, Plotly, Open3D

Best Practices

1. Standardize predictions to required format
2. Use multiple metrics for comprehensive evaluation
3. Visualize errors to understand failure modes
4. Report per-joint errors for detailed analysis
5. Cross-validate on multiple sequences

Requirements

• Python 3.8+
• NumPy, SciPy
• Matplotlib (for plotting)
• Open3D (for 3D visualization)
• PyTorch (optional, for loading models)

References

• GitHub: https://github.com/facebookresearch/hand_tracking_toolkit
• HOT3D: https://facebookresearch.github.io/hot3d
• Challenge: https://eval.ai/web/challenges/challenge-page/...

License

Apache 2.0