Theses Doctoral

Correlating Visual Speaker Gestures with Measures of Audience Engagement to Aid Video Browsing

Zhang, John

In this thesis, we argue that in the domains of educational lectures and political debates, speaker gestures can be a source of semantic cues for video browsing. We hypothesize that certain human gestures, which can be automatically identified through techniques of computer vision, can convey significant information that are correlated to audience engagement. We present a joint-angle descriptor derived from an automatic upper body pose estimation framework to train an SVM which identifies point and spread poses in extracted video frames of an instructor giving a lecture. Ground-truth is collected in the form of 2500 manually annotated frames covering 20 minutes of a video lecture. Cross validation on the ground-truth data showed classifier F-scores of 0.54 and 0.39 for point and spread poses, respectively. We also derive an attribute for gestures which measures the angular variance of the arm movements from this system (analogous to arm waving). We present a method for tracking hands which succeeds even when left and right hands are clasping and occluding each other. We evaluate on a ground-truth dataset of 698 images with 1301 annotated left and right hands, mostly clasped. Our method performs better than baseline on recall (0.66 vs. 0.53) without sacrificing precision (0.65 for both) toward the goal of recognizing clasped hands. For tracking, it results in an improvement over a baseline method with an F-score of 0.59 vs. 0.48. From this, we are able to derive hand motion-based gesture attributes such as velocity, direction change and extremal pose. In ground-truth studies, we manually annotate and analyze the gestures of two instructors, each in a 75-minute computer science lecture using a 14-bit pose vector. We observe "pedagogical" gestures of punctuation and encouragement in addition to traditional classes of gestures such as deictic and metaphoric. We also introduce a tool to facilitate the manual annotations of gestures in video and present results on their frequencies and co-occurrences. In particular, we find that 5 poses represent 80% of the variation in the annotated ground truth. We demonstrate a correlation between the angular variance of arm movements and the presence of those conjunctions that are used to contrast connected clauses ("but", "neither", etc.) in the accompanying speech. We do this by training an AdaBoost-based binary classifier using decision trees as weak learners. On a ground-truth database of 4243 video clips totaling 3.83 hours, each with subtitles, training on sets of conjunctions indicating contrast produces classifiers capable of achieving 55% accuracy on a balanced test set. We study two different presentation methods: an attribute graph which shows a normalized measure of the visual attributes across an entire video, as well as emphasized subtitles, where individual words are emphasized (resized) based on their accompanying gestures. Results from 12 subjects show supportive ratings given for the browsing aids in the task of providing keywords for video under time constraints. Subjects' keywords are also compared to independent ground-truth, resulting in precisions from 0.50-0.55, even when given less than half real time to view the video. We demonstrate a correlation between gesture attributes and a rigorous method of measuring audience engagement: electroencephalography (EEG). Our 20 subjects watch 61 minutes of video of the 2012 U.S. Presidential Debates while under observation through EEG. After discarding corrupted recordings, we retain 47 minutes worth of EEG data for each subject. The subjects are examined in aggregate and in subgroups according to gender and political affiliation. We find statistically significant correlations between gesture attributes (particularly extremal pose) and our feature of engagement derived from EEG. For all subjects watching all videos, we see a statistically significant correlation between gesture and engagement with a Spearman rank correlation of rho = 0.098 with p < 0.05, Bonferroni corrected. For some stratifications, correlations reach as high as rho = 0.297. From these results, we conclude what gestures can be used to measure engagement.


  • thumnail for Zhang_columbia_0054D_11689.pdf Zhang_columbia_0054D_11689.pdf application/pdf 11.8 MB Download File

More About This Work

Academic Units
Computer Science
Thesis Advisors
Kender, John
Ph.D., Columbia University
Published Here
November 7, 2013