Making Machine Learning Accessible for Developmental Science: The Case of Automated Face Detection

Area of Science:

• Developmental Science
• Computer Vision
• Machine Learning

Background:

• Machine learning (ML) has rapidly advanced, yet developmental science lags in adoption.
• ML offers potential for scalable, data-driven insights into developmental processes.
• Barriers to ML adoption include algorithm selection and technical implementation.

Purpose of the Study:

• To systematically evaluate state-of-the-art face detection algorithms for developmental science applications.
• To address challenges in algorithm selection and technical implementation for ML in developmental research.
• To identify high-performing face detection methods for analyzing children's egocentric visual experiences.

Main Methods:

• Evaluated 13 state-of-the-art face detection algorithms from the DeepFace library.
• Utilized data from head-mounted eye-tracking (N=20, 4- & 8-month-olds) and head-mounted cameras (N=10, 18-29-month-olds).
• Benchmarked algorithm performance against manual annotations for precision and recall.

Main Results:

• YOLOv11Face (M) and RetinaFace demonstrated superior performance, outperforming other algorithms.
• These top algorithms showed strong concordance with manual ratings, low error, and reduced systematic deviation.
• Robust rank-order correlations between algorithm and manual annotations were observed.

Conclusions:

• YOLOv11Face (M) and RetinaFace are highly effective for automated face detection in developmental datasets.
• An accessible tool, TinyExplorer Detection App, is introduced to promote ML adoption in developmental science.
• Widening access to ML tools like face detection will enhance efficiency, scalability, and innovation in the field.