Abstract
In our recent face recognition study (Chuk et al., 2013), we recruited Asian participants and used a hidden Markov model (HMM) to represent each individual's eye movement patterns. The HMM estimates regions of interests (ROIs) on the face, and the probability of transitions between ROIs. We then clustered the individuals' HMMs into two groups using a data-driven algorithm, and discovered that one group exhibited holistic eye movement patterns while the other exhibited analytic patterns. However, previous studies (Kelly et al., 2011) considered these two eye movement patterns to be markers of Asians and Caucasians, respectively. Here we recruited 24 Caucasian and 24 Asian participants to study 28 faces and then recognize them among 56 faces; eye movements were recorded. We trained one HMM per individual using all fixations, and then clustered the HMMs into two groups. We discovered that more Asians (19) than Caucasians (14) were in the analytic group. However, when the HMMs were clustered into three groups, we discovered that some Asians used a different analytic pattern that mainly shuffled between the face center and the right eye. The two races differed in their group distributions (χ2(2) = 8.064, p = .018). Since past studies suggested that the first few fixations suffice for face recognition (Hsiao & Cottrell, 2008), we also trained the individuals' HMMs using the first three fixations in each trial. Similar to the all fixation case, our clustering algorithm found two analytic and one holistic groups, but the difference in group distribution between the two races were non-significant (χ2(2) = 2.411, p = .300). In conclusion, our data-driven analyses discovered a previously unknown eye movement pattern among Asians and that cultural difference emerges after the first few fixations. These findings were not possible using previous methods that do not consider individual differences and transition information.
Meeting abstract presented at VSS 2014