Understanding First Impressions from Face Images
This project investigated how first impressions are expressed in faces through the automated recognition of these social traits from face images. First impressions are universally understood to have three dimensions: trustworthiness, attractiveness and competence. All three of these social traits have strong implications for evolution. We form first impressions of sometime within 100ms of meeting them, which is less time than it takes to blink. The first impressions we form are composed of judgements of these three social traits. This project aimed to answer the question: “Can automated methods accurately recognize social traits from images of faces?” Two main challenges are associated with the above research question, and both are related to the data available for the problem. First impressions are not something that can be objectively measured, at this point in time, and so labels for first impressions come from subjective human ratings of faces. Additionally, only one dataset is available for this problem and it has roughly 1,100 images. Training a state-of-the-art machine learning algorithm to accurately and reliably recognize social traits from face images using only 1,100 images each labeled with subjective human ratings is nearly impossible.
Student involvement: Two REU participants investigated a method for combining multiple datasets from different domains with partial labels for the problem of social trait recognition. Using this combined dataset, the students were able to train a model capable of recognizing social traits with improved accuracy over the baseline approach.
- REU Site: Collaborative Human-Robot Interaction, National Science Foundation PI: David Feil-Seifer, Co-PI: Shamik Sengupta, Amount: $360,000, Feb. 1, 2018 - Jan. 31, 2022