Exploration of factors affecting webcam-based automated gaze coding

Area of Science:

• Behavioral research
• Developmental psychology
• Human-computer interaction

Background:

• Online experiments offer benefits like larger sample sizes and diverse populations for behavioral research.
• Gaze behavior is a key outcome measure in infant studies, but manual coding is labor-intensive.
• Automatic gaze classification from webcam data could streamline infant research but faces challenges from uncontrolled environments.

Purpose of the Study:

• To create a controlled adult webcam dataset replicating noise factors from infant online studies.
• To evaluate the impact of specific noise factors on automated face and gaze detection algorithms.
• To assess the feasibility of facial anonymization for data sharing in online behavioral research.

Main Methods:

• Systematically varied factors: left-right offset, camera distance, facial rotation, and lighting direction.
• Utilized two state-of-the-art classification algorithms (iCatcher+ and OWLET).
• Tested the effect of morphing faces to be unidentifiable on algorithm performance.

Main Results:

• Facial detection was most sensitive to lighting conditions.
• Gaze coding accuracy was consistently impacted by camera distance and lighting.
• Facial anonymization did not significantly degrade performance, suggesting its utility for data transparency.

Conclusions:

• Camera distance and lighting are critical factors affecting gaze coding accuracy in online studies.
• Facial anonymization is a viable option for sharing online behavioral data.
• The developed dataset and findings will enhance the robustness of online testing for developmental psychology research.