In Experiment 3, we compare search performance for face average and multiple images search templates. We also investigate the impact of set size on performance, by varying the number of distractors in the array. This enables us to estimate the relationship between the set size and performance, which is an important consideration when estimating the accuracy of visual search in real-world settings. By estimating this “search-slope” function, it is possible to estimate how speed and accuracy would be impacted by larger crowd sizes (Tong & Nakayama, 1999; Treisman & Souther, 1985).
Method
Participants
Twenty-six undergraduate students (17 women, mean age 19.2 years, SD = 1.9) took part in the experiment in exchange for course credit. All participants reported normal (or correct-to-normal) visual acuity and normal color vision. Informed consent was obtained before the experiment.
Materials and procedure
Experiment 3 employed the same materials and procedure used in Experiments 1 and 2 except that we varied the search Set Size (either 5, 10, or 20 images). Each target identity appeared three times (once at each set size), randomly located in the array along with 4, 9, or 19 fillers. The assignment of search templates to set size conditions was randomized across stimulus identities and counterbalanced across participants. In this experiment, an average of 13 of the 20 international celebrities (SD = 4) were correctly identified. None of the Dutch celebrities were familiar to any of the participants.
Results
Two 2 × 2 × 3 repeated measures ANOVAs were used to analyze the accuracy and response time of participant’s responses with Familiarity (unfamiliar, familiar), Template (average, multiple), and Set Size (5, 10, 20) as the factors (Fig. 3).
Accuracy
For accuracy, there was a significant main effect of all three factors. First, there was a main effect of Familiarity, F(1, 25) = 274.27, p < 0.001, ηp2 = 0.916, with accuracy for familiar faces being higher than for unfamiliar faces (familiar = 87.9%, unfamiliar = 61.5%). There was also a main effect of Template, F(1, 25) = 17.45, p < 0.001, ηp2 = 0.411, with the multiple image template having higher accuracy than the average template (multiple = 77.3%, average = 72.1%). The main effect of Set Size was also significant, with increasing Set Size causing a decrease in accuracy, F(2,50) = 136.34, p < 0.001, ηp2 = 0.845 (Set Size 5 = 86.8%, Set Size 10 = 76.2%, Set Size 20 = 61%). There was a significant interaction between the Familiarity and Template factors, F(1, 25) = 22.94, p < 0.001, ηp2 = 0.478. Simple main effects analysis reveal that multiple image template has significantly higher accuracy than the average template for unfamiliar faces, F(1, 25) = 43.03, p < 0.001, ηp2 = 0.633, but that there was no difference between Templates for familiar faces, F(1, 25) = 0.22, p = 0.641, ηp2 = 0.009. A significant interaction was also observed between Familiarity and Set Size, F(2, 50) = 8.99, p < 0.001, ηp2 = 0.265. Based on Fig. 3, it appears that there is a shallower search slope (i.e. smaller cost of larger set size) for familiar than unfamiliar faces. A formal analysis of search slopes is presented below. The three-way interaction, F(2, 54) = 0.56, p = 0.58, ηp2 = 0.022, and the interaction between Template and Set Size was not significant, F(2, 54) = 1.50, p = 0.234, ηp2 = 0.056.
Response time
For response time, there was an overall main effect of Familiarity, F(1, 25) = 9.77, p = 0.004, ηp2 = 0.281, with familiar face search faster than unfamiliar face search (familiar = 3.4 s, unfamiliar = 3.9 s). There was also a main effect of Template, F(1, 25) = 31.07, p < 0.001, ηp2 = 0.554, with multiple image templates resulting in faster search time than average image template (multiple = 3.4 s, average = 3.9 s). We also found a significant main effect for Set Size, F(2, 50) = 206.75, p < 0.001, ηp2 = 0.892, with increasing Set Size resulting in longer response time (Set Size 5 = 2.4 s, Set Size 10 = 3.5 s, Set Size 20 = 5 s). We also found no significant three way interaction, F(2, 50) = 1.70, p = 0.193, ηp2 = 0.064, nor two way interactions between Familiarity and Template, F(1, 25) = 2.48, p = 0.128, ηp2 = 0.090, Familiarity and Set Size, F(2, 50) = 0.24, p = 0.784, ηp2 = 0.010, or Template and Set Size, F(2, 50) = 1.39, p = 0.260, ηp2 = 0.052.
Search slopes
To compare the speed and accuracy that distractors are rejected, we also calculated and compared the search slopes for each condition (Tong & Nakayama, 1999; Treisman & Souther, 1985). Search slopes were calculated for both accuracy and response time function and analyzed in a 2 × 2 repeated measures ANOVA with Familiarity and Template as the factors. Slopes were calculated for each participant individually using linear modelling by the least squares method, with the gradient of the model giving the measured search slope.
For accuracy, we found a significant main effect of Familiarity, F(1, 25) = 5.5, p = 0.007, ηp2 = 0.255, with the slope for familiar faces being significantly shallower than for unfamiliar faces, suggesting that familiarity reduced the detrimental impact a larger search array had on P’s ability to find the target. The main effect of Template, F(1, 25) = 0.09, p = 0.767, ηp2 = 0.004, and interaction was not significant, F(1, 25) = 0.88, p = 0.357, ηp2 = 0.034.
For response time, we found no significant main effect of Familiarity, F(1, 25) = 0.32, p = 0.577, ηp2 = 0.019, or Template, F(1, 25) = 2.03, p = 0.166, ηp2 = 0.075, nor interaction between these two factors, F(1, 25) = 0.59, p = 0.450, ηp2 = 0.023, confirming that these factors did not interact with the response time search slope.
Discussion
In Experiment 3, we have found that search templates elicited from multiple images lead to faster and more accurate search than those from face averages. These results extend the findings of studies of unfamiliar face matching (White et al., 2014), where multiple images have been shown to lead to more accurate matching performance than face averages. Although face averages contain the invariant features of a face, it appears that the variance information contained in multiple images produces more robust templates for visual search.
Moreover, the advantage for multiple images occurred despite substantial differences in the quantity of image information used to derive these representations: multiple image arrays consisted of four images and face averages were generated from 19 images. This discrepancy may mask a larger advantage for multiple images than reported here, as combining more images into the face averages may improve subsequent identifications (Burton et al., 2005). Given that access to images of targets may be limited in real-world tasks, this suggests that face averages are of limited applied use for this task.
Another aim of this experiment was to determine whether the number of distractors affected changes in familiarity and the search template. Our results show that performance declines precipitously with increasing number of distractors. However, for accuracy scores at least, this effect is substantially larger for unfamiliar faces than for familiar faces. Because neither familiarity nor multiple image exposure provide additional benefits to response time as the number of distractors increases, this suggests that these familiarity-based improvements are not a result of faster distractor rejection. Because face processing is capacity-limited, with only one face being processed at a time (Bindemann, Burton, & Jenkins, 2005; Bindemann, Jenkins, & Burton, 2007), these findings support the conclusion that searching for a particular face, whether familiar or unfamiliar, must be performed serially (Nothdurft, 1993; Tong & Nakayama, 1999; Wolfe & Horowitz, 2004).
Overall, the results of Experiment 3 show that becoming familiar with a face can help protect against costly false-positive errors when searching for faces in crowds and that partial benefits of familiarity can be reached by exposing participants to image sets that have naturalistic variation in facial appearance.