Automatic pose estimation in newborn infants: Lessons from the Baby Grow study

Area of Science:

• Computational neuroscience
• Developmental psychology
• Biomedical engineering

Background:

• Machine learning advances enable analysis of infant motor skills in natural settings.
• Evaluating pose estimation algorithms is crucial for reliable infant movement analysis.

Purpose of the Study:

• To assess the performance of state-of-the-art pose estimation models in home-based infant video analysis.
• To identify strengths, limitations, and optimal models for real-world infant movement research.

Main Methods:

• Analyzed 2,640 frames from 22 home videos of 8 newborns (2, 4, 8 weeks old).
• Tested MediaPipe, OpenPose, PCT, RTMpose, Sapiens, and VitPose using varied video conditions (clothing, lighting, angles).
• Evaluated models based on accuracy (OKS, PCKh) and processing speed.

Main Results:

• RTMpose achieved the highest accuracy; MediaPipe demonstrated the fastest processing speed.
• MediaPipe showed practical utility by balancing speed and accuracy.
• RTMpose, Sapiens, and VitPose proved most robust across diverse environmental conditions.

Conclusions:

• Newer pose estimation models surpass legacy tools like OpenPose for infant movement analysis.
• Video context and model selection critically impact pose estimation accuracy in real-world settings.