CI-SUSTAIN: Collaborative Research: Extending a Large Multimodal Corpus of Spontaneous Behavior for Automated Emotion Analysis

点击翻译按钮获取中文摘要

This project will extend and sustain a widely-used data infrastructure for studying human emotion, hosted at the lead investigator's university and available to the research community. The first two versions of the dataset (BP4D and BP4D+) contain videos of people reacting to varied emotion-eliciting situations, their self-reported emotion, and expert annotations of their facial expression. Version 1, BP4D (n=41), has been used by over 100 research groups and supported a successful community competition around recognizing emotion. The second version (BP4D+) adds participants (n = 140), thermal imaging, and measures of peripheral physiology. The current project greatly broadens and extends this corpus to produce a new dataset (BP4D++) that enables deep-learning approaches, increases generalizability, and builds research infrastructure and community in computer and behavioral science. The collaborators will (1) increase participant diversity; 2) add videos of pairs of people interacting to the current mix of individual and interviewer-mediated video; 3) increase the number of participants to meet the demands of recent advances in "big data" approaches to machine learning; and 4) expand the size and scope of annotations in the videos. They will also involve the community through an oversight and coordinating consortium that includes researchers in computer vision, biometrics, robotics, and cognitive and behavioral science. The consortium will be composed of special interest groups that focus on various aspects of the corpus, including groups responsible for completing the needed annotations, generating meta-data, and expanding the database application scope. Having an infrastructure to support emotion recognition research matters because computer systems that interact with people (such as phone assistants or characters in virtual reality environments) will be more useful if they react appropriately to what people are doing, thinking, and feeling. The team will triple the number of participants in the combined corpora to 540. They will develop a dyadic interaction task and capture data from 100 interacting dyads to support dynamic modeling of interpersonal influence across expressive behavior and physiology, as well as analysis of emotional synchrony. They will increase the density of facial annotations to about 15 million frames in total, allowing the database to become sufficiently large to support deep-learning approaches to multimodal emotion detection. These annotations will be accomplished through a hybrid approach that combines expert coding using the Facial Action Coding System, automated face analysis, and crowdsourcing with expert input from the research community. Finally, the recorded data will be augmented with a wide range of meta-data derived from 2D videos, 3D videos, thermal videos, and physiological signals. To ensure the community is involved in sustaining the infrastructure, in addition to the governance consortium described above, the investigators will involve the community in jointly building both APIs that allow adding meta-data and annotations and tools to support the submission and evaluation of new recognition algorithms, then organizing community-wide competitions using those tools. The research team will also reach out to new research communities around health computing, biometrics, and affective computing to widen the utility of the enhanced infrastructure, grow the community of expert annotators through training workshops, and build an educational community around the infrastructure that facilitates the development and sharing of course materials that use it. Long-term, the infrastructure will be funded through a combination of commercial licensing and support from the lead university's system administration group.