Which faces

But that skill could become obsolete in three years or so, when machine learning advances enough to iron out the imperfections. Even now, computer-generated people pose a problem.

Contact Jane Lytvynenko at jane. Got a confidential tip? Submit it here. Subscribe to get the best Verge-approved tech deals of the week. Cookie banner We use cookies and other tracking technologies to improve your browsing experience on our site, show personalized content and targeted ads, analyze site traffic, and understand where our audiences come from.

By choosing I Accept , you consent to our use of cookies and other tracking technologies. Cybersecurity Mobile Policy Privacy Scooters.

Phones Laptops Headphones Cameras. Tablets Smartwatches Speakers Drones. Accessories Buying Guides How-tos Deals. Health Energy Environment. YouTube Instagram Adobe. Kickstarter Tumblr Art Club. Film TV Games. Fortnite Game of Thrones Books. Comics Music. Filed under: Tech Artificial Intelligence.

All of these scenarios are significantly impeded when actors are wearing masks. To overcome, at least partially, the devastating consequence of masks, and to be able to offer remedies, we need to first understand how face masks interfere with basic aspects of face recognition. What primary dimensions are most vulnerable to masks? And are there certain contexts under which people can circumvent the detrimental effects of masks? The vast literature on face masks addresses a broad scope of issues.

Here we focus on studies that tested for the effects of masks on basic process of face recognition, in particular the perception of identity and emotion. Research dating from the pre COVID era with masked and occluded faces has revealed a reduction in recognition of facial identity Dhamecha et al.

In one study Stephan and Caine, recognition accuracy of familiar and unfamiliar faces declined when different features of the face e. In another study Fitousi and Wenger, , recognition of identity was hampered in both accuracy and speed when sunglasses were added.

Similarly, categorization of facial expression was disrupted when a scarf covered the mouth area see also, Kret and De Gelder, ; Noyes et al.

In one of the first COVID studies by Carragher and Hancock participants judged whether two simultaneously presented faces showed the same personal identity or two different identities. Surgical masks had a large detrimental effect on face matching performance irrespective of whether one or two of the faces were masked.

The impairment was similar in size for familiar and unfamiliar faces. Other studies have shown that faces reduce emotion recognition accuracy Carbon, ; Grundmann et al.

In one important study, Carbon presented faces with six different emotional expressions in either a fully visible condition or in a partly covered with masks condition. He found lower accuracy and confidence levels in the masked faces condition. In addition, observers misinterpreted disgusted faces as being angry, while other emotions e.

All of these studies suggest that face masks interfere with basic mechanisms of face recognition to some degree. But what are those mechanisms that are responsible for the impairment? Previous studies Carragher and Hancock, ; Freud et al. While most of these efforts pointed to the dramatic effects of masks on these aspects, many practical and theoretical issues have remained unanswered: a the degree to which masks affect speed of processing of faces, b the influence of masks on other primary facial dimensions such as age, and gender which play a central role in social cognition Darwin, ; Freeman et al.

All of these bear tremendous import for more advanced processes that depend on correct and efficient initial person construal processes Freeman and Ambady, ; Fitousi, b , such as impression formation and stereotyping Oosterhof and Todorov, The recent literature on face masks may lead us to think that face masks universally distort the recognition of all types of facial dimensions e.

However, it is imminently possible that the influence of masks is selective, with some facial attributes being immune to such influences. It might be the case that certain facial dimensions can be efficiently extracted from face areas that are not occluded by the mask. Carbon has shown that recognition of fear in contrast to other emotions was not hampered by masks, probably because fear is decoded from the top part of the face.

Similarly, Kastendieck et al. So how can one predict what face attributes are going to be distorted by masks? Using this technique Blais et al. It can inform us on the importance of the mouth area in the categorization of various face dimensions. A related issue that deserves a comment concerns the deployment of strategies.

It is also plausible that observers have developed specific strategies during the COVID period to circumvent the obstacles posed by masks. In natural everyday circumstances we can still interact with other people, as well as perceive their identity, gender, age, emotion and other social aspects, in spite of the debilitating conditions of masks.

This is also true with respect to other sub-optimal conditions under which we perceive faces such as: poor lighting, low acuity, deformation in shape, in which we can still recognize faces quite efficiently. People have learned to cope with such debilitating conditions by harnessing unique processing strategies. For example, people can take advantage of additional contextual cues from voice Golan and Baron-Cohen, ; Knoblauch et al.

Moreover, not all aspects of performance e. For example it might be the case that people trade off their accuracy for speed, or alternatively, compromise their speed to achieve high levels of accuracy.

Research programs should allow us to better understand the relations between speed and accuracy and the circumstances under which such strategies may arise. A plausible candidate for the detrimental effects of masks on face recognition is the disruption of a psychological mechanism known as holistic processing Young et al.

According to this idea, faces are perceived as wholes, and the constituting parts or features are secondary to the global Gestalt but see, Fitousi, , , , a. Disruption of holistic processing mechanisms through occlusion of face parts may lead to an impairment in face recognition.

Researchers Maurer et al. Face masks cover areas of the mouth and nose, and therefore might hamper any or all of these types of representations. In a large-scale online study, Freud et al. As predicted, they found considerable decrease in performance in this test with masked faces. Interestingly, the face inversion effect FIE, Yin, —impaired performance with inverted faces—was smaller with masked compared to unmasked faces.

The FIE Prete et al. This argument is supported by other studies showing that holistic processing is important for normal face perception abilities, and that it is disproportionally contributed by the lower mouth region part of the face compared with the upper eye region part Tanaka et al. However, it should be noted that some researchers question the view that the FIE is a marker of holistic processing Sekuler et al.

Another source of evidence for the role of the lower part of the face in holistic face recognition comes from the composite face effect CFE, Young et al.

In this paradigm, participants are presented with a face that is composed from top- and bottom- halves of two different identities. This composite face creates the illusion of a newly, never-seen-before face. There are several versions of paradigm Fitousi, , but the most used task is often to judge the top part of the face and ignore its bottom part.

The typical result shows that when the two face halves are aligned, participants cannot ignore the identity of the bottom part, this effect is reduced or abolished when two face halves are misaligned.

The CFE has been often taken as evidence for holistic processing of faces. The CFE illusion is typically observed in tasks that require discrimination between two identities, but it has been also documented in tasks that call for emotion Calder et al. In these versions, for example, people tend to perceive the top half of the face as less masculine if the lower part is more feminine.

This illusion gives currency to the hypothesis that the lower part of the face is essential for the categorization of identity, emotion, gender, and age. However it should be noted that not all researchers hold to the view that faces are processed holistically Tversky and Krantz, ; Massaro and Friedman, ; Donnelly et al.

The current study addresses various theoretical and practical issues. First, we were interested in the impact of face masks on the speed and accuracy of categorization of primary facial dimensions. Second, we wanted to uncover the mechanisms governing potential disruptions to normal processing. Third, we tested for the presence of unique processing strategies.

All of these issues are addressed here in a systematic and comprehensive fashion. The face stimuli consisted of realistic images created in our lab by photographing volunteers with and without masks.

Each experiment focused on one facial dimension e. We also tested for the FIE, with face images appearing either upright or inverted. In addition, we probed the existence of processing strategies using blocked and mixed designs. Our hypotheses were as follows: a given the importance of the lower part of the face in many aspects of face categorization, we hypothesized that face masks will disrupt both accuracy and speed in all types of categorizations identity, emotion, gender, and age with upright-faces.

If indeed, the source of this effect is disruption of holistic processing Freud et al. In contrast, if holistic processing does not play a role in face recognition Fitousi, , then inversion should not interact with the effect of face masks. In addition, we predicted that c if observers are capable of developing strategies for coping with the unwarranted effects of face masks, they should be more likely to do so in blocked- compared to mixed designs.

This is because the regularity of the faces in this type of blocks should allow participants to exert top-down control and mitigate the unwarranted effects of masks. A total of participants took part in Experiments 1—4. They were recruited from Ariel University pool of participants and compensated with a course credit.

All observers had normal or corrected-to-normal vision. The faces in the four experiments were realistic gray-scale images created in our lab. This is important because the great majority of previous studies have not taken pictures with real people wearing masks, but have artificially patched an image of a generic mask on images of unmasked faces Carragher and Hancock, ; Freud et al.

This practice is not ecologically valid because real masks can convey valuable structural information via curvature, depth, light-shading, and other types of cues. This information is lost when the masks are added with a photo editing software to an originally unmasked face. Our images conveyed front-views of real people of Caucasian ethnicity, who were photographed with and without a surgical mask covering the lower part of their face including the mouth and the nose see Figure 1.

The faces in our study varied on four primary dimensions: a identity, b emotion, c age and d gender. There is a broad consensus among face researchers that these dimensions are the most important for social interaction Freeman and Ambady, ; Fitousi, b. To create the face stimuli, we recruited 16 volunteers from Ariel community area who agreed to take part in a photo-shooting session.

All volunteers provided written consents in which they gave their permission to use the pictures in experiments and journal publications.

The 16 volunteers consisted of four young females, four young males, four old females, and four old males. The young actors were in their twenties whereas the old actors were in their sixties or seventies. The same photographer took four front-view photos of each volunteer: a neutral expression without a face mask, b neutral expression with a face mask, c angry expression without a face mask, and d angry expression with a face mask. All images were taken as color pictures under identical position and light conditions.

The color images were then converted into gray-scale images and standardized with a commercial photo editing software. We then elected the best images of 8 identities four males and four females out of the 16 identities. This resulted in 32 images. The faces varied on four primary facial dimensions: age, gender, emotion, and identity. This allowed us a tight control over the stimuli set. Figure 1. Examples of face images used in the four experiments. The top row presents an example for a young female with neutral emotion.

The bottom row gives and example for an old female with angry emotional expression. The exact same set of faces was deployed in all four experiments. The experiments were also identical with respect to the procedure and design. They only differed with respect to the target facial dimension to which the participants were asked to pay attention. That is, on each experiment, participants categorized the faces on a single predefined facial dimension e.

The task was a speeded two-choice categorization. Participants pressed on one of two keys to indicate their decision. A different group of participants was recruited for each experiment.

It should be noted that none of the participants was familiar with the identities of the faces in the experiments.

This was confirmed before the experiment took place. In this sense, the faces were unfamiliar. In Experiment 1, participants categorized the faces according to the face's identity. We grouped the 8 identities in the stimuli set into two artificial social groups A vs. We kept the values of the other three facial dimensions age, gender, emotion equally distributed across these two groups. On each trial, participants were presented with a face from one of the two groups and had to decide, while timed, whether that face belonged to group A or B.

To reduce the potential influence of memory load, we presented small images of the faces of each group on the bottom of the screen on each trial. Participants could therefore use these images to memorize the correct grouping. Because the faces were arbitrarily assigned to each group, the groups had no unique characterizations, and were described to the participants as simply group A and group B.

The task does not require familiarity with the identities or deep semantic analysis. In this sense, the task is comparable to the traditional 'same-different' task in which participants judge whether two simultaneously presented face photographs show the same person or two different people Carragher and Hancock, In Experiment 2, participants classified the faces according to the face's emotion neutral vs. In Experiment 3, the target dimension was age young vs.

In Experiment 4, participant categorized the faces according to gender man vs. Each experiment consisted of 6 blocks: Block 1 —upright faces with half of the faces wearing face masks and half of the faces without masks, Block 2 —upright faces with all faces wearing masks, Block 3 —upright faces with no masks, Block 4 —inverted faces with all faces wearing masks, Block 5 —inverted faces with no masks, and Block 6 —inverted faces with half of the faces wearing face masks and half of the faces not wearing masks.

In combination, the six experimental blocks created two nested experimental designs. Blocks 2 — 5 formulated a blocked design in which the factors of Mask mask, no mask and Inversion upright, inverted were manipulated across blocks. Blocks 1 and 6 created a mixed design, with the factor of Mask mask, no mask being manipulated within a block, while the factor of Inversion upright, inverted being manipulated across blocks. Both designs were administrated within-subject with repeated measures.

The order of blocks was random. We purposely elected to use these two types of designs to test whether participants will develop specific strategies to cope with the potentially delimiting influence of face masks. Such strategies if exist , should most likely surface in the blocked design, but not in the mixed design. We hypothesized that the results in the two designs should be identical if participants do not develop unique strategies. All the faces were presented in an upright position.

Block 6 which was also part of the mixed design, consisted of the same 32 faces from Block 1 , but now in an inverted position. Blocks 2 — 5 were part of the blocked design. Each of them included 16 trials, which incorporated half of the trials from Block 1. The frequency of dimensional levels was balanced also in these blocks, with equal number of identities, gender, and emotion per block. Block 2 consisted of only masked faces in an upright position. Block 3 presented only faces with no masks in an upright position.

Block 4 consisted of only masked faces in inverted position. And Block 5 presented only faces with no masks in an inverted position. Ten trials of training preceded each block. All the experiments were run with the Macromedia Authorware software Macromedia, Participants used the index fingers of the left and right hand, respectively. After responding, the image was removed from the screen, and the screen remained empty for ms.

In Experiment 1 participants were asked to categorize eight identities into two arbitrary social groups. The experiments were conducted online.