Online testing of infants by recording video with a webcam has the potential to improve the replicability of developmental studies by facilitating larger sample sizes and by allowing methods (including recruitment) to be specified in code. However, the recorded video still needs to be manually scored. This labour-intensive process puts downward pressure on sample sizes and requires subjective judgements that may not be reproducible in a different laboratory. Here we present the first fully automatic pipeline, using a face analysis software-as-a-service and a discriminant-analysis classifier to score infant videos acquired online. We compare human and machine performance for looking time and preferential looking paradigms; machine performance demonstrates a promising proof of principle for looking time and is above chance in classifying preferential looking. Additionally, we studied the characteristics of the video and the child that influenced automated scoring, so that future studies can acquire data that maximises the performance of automatic gaze coding and/or focus on improving automatic coding for particularly challenging data. We believe this technology has great promise for developmental science.
Keywords: Face detection; Looking time; Machine vision; Preferential looking; Webcam.
Copyright © 2018. Published by Elsevier Inc.