Comparison of marker-less 2D image-based methods for infant pose estimation.

Area of Science:

• Biomedical Engineering
• Developmental Pediatrics
• Computer Vision

Background:

• Automated General Movement Assessment (GMA) relies on accurate infant pose estimation from video.
• Existing pose estimators are often generic (trained on adults) or specialized for infants, with varying performance.

Purpose of the Study:

• To compare the performance of generic and specialized infant pose estimators for automated GMA.
• To evaluate the impact of viewing angle (diagonal vs. top-down) on pose estimation accuracy in infants.

Main Methods:

• Utilized 4500 annotated video frames from 75 infant recordings (4-16 weeks).
• Computed error and Percentage of Correct Key-points (PCK) to assess pose estimation accuracy.
• Compared performance of generic models (e.g., ViTPose) and infant-specific models, and evaluated diagonal vs. top-down camera views.

Main Results:

• The generic ViTPose model, trained on adults, demonstrated superior performance on the infant dataset.
• No significant improvement was observed with specialized infant pose estimators compared to generic ones.
• Retraining a generic model on infant data significantly improved pose estimation accuracy.
• Top-down camera views yielded significantly better pose estimation accuracy than the conventional diagonal view.

Conclusions:

• Generic pose estimators, particularly when retrained on target data, are effective for infant pose estimation in GMA.
• Infant-specific pose estimators show limited generalization capabilities across different infant datasets.
• A top-down viewing angle is recommended for future automated GMA recording setups to enhance accuracy.