Poster #118 - Eye Spy: Investigating the Reliability of Coding Eye Contact from a Camera Disguised as Glasses.

• In Event: 1-061 - Poster Session 02
  In Poster Session: PS 02 Section - Methods, History, Theory

Integrative Statement

Background: Behavioral assessments targeting social communication of individuals often examine eye gaze behavior (e.g., atypical gaze patterns of individuals with autism spectrum disorder; Chawarska, Volkmar, & Klin, 2010). Gaze behavior is traditionally studied by coding videos obtained from stationary cameras. Coding footage from these cameras require rater inference to determine the location of a participant’s gaze. We aimed to examine the accuracy and reliability of Pivothead SMART™, commercially available glasses with a high-resolution point-of-view (POV) camera embedded within the bridge, as an alternative to the traditional stationary cameras in detecting gaze behavior. It was hypothesized that the POV camera technology will produce a more accurate measurement of gaze behavior via minimizing subjective inference when reviewing footage.

Methods: Each of the 20 participants were instructed to look at 20 specific locations on a mannequin head wearing the POV device, while the glasses camera and a stationary camera (i.e. “methods”) simultaneously captured the gaze behavior. Video recordings of gaze behavior of each subject were independently coded by two blind raters with the aim of classifying a given gaze as targeting four areas of interest (AOI; forehead, eyes, nose, and mouth). Following random assignment, each rater coded glasses camera recordings of 10 subjects and the stationary camera recordings of the other 10 subjects. Two additional reliability raters independently coded 35% of instances. For each method, the accuracy of the raters in classifying a subject’s gaze into one of the four AOIs was determined by comparing with the prompt given to each subject at the corresponding instance (i.e. the “truth”). Sensitivity, specificity, positive and negative predictive values (PPV and NPV), and agreement of the two raters with the reliability raters (i.e. Cohen’s ) were compared via Fisher’s exact tests. Area under the receiver operating characteristic (ROC; i.e. 1-specificity vs. sensitivity) curves of the two methods were compared using a bootstrap approach.

Results: On average, participants were 95% confident of their ability to look at a prompted AOI. Significant differences between methods (favoring the glasses camera) were found for the sensitivity of responses to the nose AOI (p < 0.05), specificity of responses to the mouth (p < 0.005), PPV for both the nose and mouth (p < 0.05; p < 0.005), and NPV for the nose (p < 0.05) (see Table 1). ROC curve comparisons yielded significant differences for the nose and mouth (p < 0.001; p < 0.05) and a trend for the eyes (p = 0.077) favoring the glasses camera (see Figure 1). Furthermore, significant differences favoring the glasses camera were found for the inter-rater agreement of the nose AOI (p < 0.005). Compared to the glasses camera, none of the findings were in support of the stationary camera.

Conclusions: Our results suggest that using a camera located on the glasses of a “target” individual seem to outperform a stationary camera in classifying gaze behavior of a subject, particularly when targeting the nose and mouth AOIs. Future studies are warranted to confirm the usefulness of this method for studying human gaze behavior.

Authors

• Blake D. West, Texas Tech University - Human Development & Family Studies
  Presenting Author

• Alexandra C. Abide, Texas Tech University
  Non-Presenting Author

• Miranda A. Cox, Texas Tech University
  Non-Presenting Author

• Georgina J. Rosenbrock, Texas Tech University
  Non-Presenting Author

• Chanaka N. Kahathuduwa, Texas Tech University
  Non-Presenting Author

• Ann Mastergeorge, Texas Tech University
  Non-Presenting Author