Gaze Classification on Redacted Videos

Videos have become prevalent, and a dominant mode of communication and consumption, providing rich content. On the other hand, stemming from privacy considerations, the process of redaction masks prominent individuals, or conversely the background, and conceals information. These two concepts takes us in opposing directions. In this work we highlight the importance of feature preserving face redaction in videos in the context of behavioral analysis. Specifically, we consider a well known non-verbal assessment method commonly used in studying the development of young children. Several pairs of social and nonsocial videos are presented on a mobile device, and using the front camera, the proportion of gaze directed to the social videos are analyzed. Can a redacted video taken by the care-giving parent sent to a server be automatically analyzed without identifying the child? We answer this question in the affirmative – gaze classification can be done with about equal, significantly high accuracy for both the redacted and the original, non-redacted video.