Title: Computational Methods for Measurement of Visual Attention from Videos towards Large-scale Behavioral Analysis

Eunji Chong

School of Computer Science

College of Computing

Georgia Institute of Technology

Date:  Thursday, January 9th, 2020

Time: 3:30 - 5:30 PM (EST)

Location: TSRB 222

Committee:

Dr. James M. Rehg (Advisor), School of Computer Science, Georgia Institute of Technology

Dr. Agata Rozga, School of Computer Science, Georgia Institute of Technology

Dr. Gregory D. Abowd, School of Computer Science, Georgia Institute of Technology

Dr. Irfan Essa, School of Computer Science, Georgia Institute of Technology

Dr. Yaser Sheikh, Robotics Institute, Carnegie Mellon University

Abstract:

Visual attention is a critically-important aspect of human social behavior, visual navigation, and interaction with the 3D environment, and where and what people are paying attention to reveals a lot of information about their social, cognitive, and affective states. While monitor-based and wearable eye trackers are widely available, they are not sufficient to support the large-scale collection of naturalistic gaze data in contexts such as face-to-face social interactions or object manipulation in 3D environments. Wearable eye trackers are burdensome to participants and bring issues of calibration, compliance, cost, and battery life.

This thesis investigates different ways to measure real-world human visual attention using computer vision from plain videos and its use for identifying meaningful social behaviors. Specifically, three methods are investigated. First, I present methods for detection of looks to camera in first-person view and its use for eye contact detection. Experimental results show that the presented method can achieve the first human expert-level detection of eye contact. Second, I develop a method for tracking heads in a 3d space for measuring attentional shifts. Lastly, I propose spatiotemporal deep neural networks for detecting time-varying attention targets in video and present its application for the detection of shared attention and joint attention. The final method achieves state-of-the-art results on different benchmark datasets on attention measurement as well as the first empirical result on clinically-relevant gaze shift classification.

Presented approaches have the benefit of linking gaze estimation to the broader tasks of action recognition and dynamic visual scene understanding, and bears potential as a useful tool for understanding attention in various contexts such as human social interactions, skill assessments, and human-robot interactions.