Vision-Based Happiness Inference: A Feasibility Case-Study

Abstract

To humanize interaction between users and computers, one needs the ability to infer the users' mood. One approach is to use a vision-based approach. We quantify the 'preview effect' bias in visual mood assessment. We demonstrate that automated tools which infer user mood from photographs or video may be affected by the presentation methodology used while performing image capture. Specifically, we demonstrate that showing a "preview" of oneself, i.e., a mirror, increases the accuracy of the visual mood inference algorithms present in Google's Mobile Vision API. Our findings show that studies that incorporate visual mood assessment should include "preview" images to reduce bias and increase th..