Project Details
Description
We propose to build upon nearly 5 years of work on the Socio-Cultural Attitudinal Networks MURI grant by performing the following tasks that leverage the Resistance game dataset collected during prior years of the SCAN project. The principal goal of our work during the time frame envisaged on the SCAN project involves learning models of trust from videos of group interactions using nonverbal communications alone.
When US soldiers encounter a group of individuals, especially in situations where they do not have command of the native language and culture involved, they need to have the ability to dynamically learn how much one person A in the group trusts another person B in the group. The trust-distrust relationships between people play an important role in the final decisions made by the group. In the absence of a detailed understanding of the groups’ native language and culture, US personnel must use non-verbal cues such as facial expressions, emotions, gestures, and vocal cues. We promise to look at two problems:
1. Given a video clip of a group G of subjects, can we predict which person is the most trusted person (MTP) in the group?
2. Given a video clip of a group G of subjects and two subjects A and B, can we solve the bilateral trust problem (BTP), i.e. how much A trusts B on a 1-5 scale?
In order to achieve these tasks, we will propose an ensemble architecture that uses several novel features:
i) The principal novelty of our work will be based on defining a family of novel mutual interaction graphs MIG(t). For each small time window t (perhaps consisting of 10 seconds of video) MIG(t) will capture the mutual interactions between people, e.g. if person A is looking at person B during the time window t, then there will be an edge from A to B labeled with the fact that A is looking at B for w% of the time window t. A number of other types of mutual interactions are also possible – a variant of this definition would place an edge between A and B if the
Status | Finished |
---|---|
Effective start/end date | 9/1/21 → 11/30/22 |
Funding
- University of Maryland, College Park (104773-Z8424106-B // W911NF1610342)
- Army Research Office (104773-Z8424106-B // W911NF1610342)
Fingerprint
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.