Human face detection in natural scenes

Lead Research Organisation: University of Glasgow
Department Name: School of Psychology

Abstract

Humans are particularly good at finding a face in a complex visual scene, but to date there is very little known about how this can be achieved. In contrast, there is a large body of evidence describing research on facial identity, expression, attractiveness and so forth.  However, in order to make decisions about any of these dimensions, one must first find a face, and focus one’s attention upon it. In this sense, the detection of faces underlies all these other processes, and yet it is very poorly understood.  


The research in this project will examine face detection in real scenes. A database of images will be constructed, in which faces are shown in a variety of viewpoints, are occluded to various extents, vary in size, luminance and colour. Viewers will be shown these images across a series of experiments. The aim is to establish which features of a face ‘trigger’ its detection in a visual scene. This will increase the understanding of fundamental processes with underlie our social perception.

Publications

10 25 50

publication icon
Bindemann M (2009) The role of color in human face detection. in Cognitive science

publication icon
Burton AM (2009) The role of view in human face detection. in Vision research

 
Description In Strand A we performed a sequence of ten experiments. The main results of these studies can be summarized as follows. First, viewers are very fast and accurate to detect a face in a scene, though faces in profile are harder to detect than those in full-face or three-quarter views. Second, the advantage for full-face and three-quarter views is not due to the fact that users are detecting ‘a pair of eyes’. When faces are cut in half, a full-face view (now showing a single eye)_is still detected just as fast and accurately as a face rotated to mid-profile and displaying both eyes. Both these views are still detected more quickly than a profile. Third, colour helps viewers to detect faces. Furthermore, colour seems to be used in an intelligent way by the visual system, because it is only useful when it conforms to skin-tone, and retains the appropriate shape. Fourth, although viewers are quick to spot a face in a scene, their first eye fixation tends not to land on the face. This is important, because some previous research has suggested that faces somehow pull the attention of viewers towards them. In fact, when controlled experiments are performed, it is clear that people tend to look at the centre of any scene first, and only then at the face. In fact, the upper body receives more initial eye-movements than the face when viewing a person in a scene – a result which is novel to this research programme.

In Strand B we performed a further 7 experiments. Here we were particularly interested in which parts of a face tend to receive a viewer’s attention, and so we presented faces in isolation. Previous research has tended to stress the importance of the eyes and the centre of the face. However, early research tended to present front-view faces in the centre of a display, and so the centre of the face is confounded with the centre of the display. In these experiments we discovered that viewers do not tend to look at the eyes first. Instead, they tend to look first at the centre of the head, and only then towards the eyes. When the view is such that the centre of the head is near the eyes (full face), people do tend to look at the eyes first. However, when the centre of the head is a cheek (as it is in profile view) then people look at the cheek first. Once again, this finding extends previous research by making it clear that the eyes do not automatically ‘pull’ one’s attention towards them. People do tend to look at the eyes, but this does not happen in the initial eye-movement, and is not mandatory. We went on to explore this further using an ‘attentional-cueing’ procedure. Once again, we demonstrated that people do tend to focus on person’s eyes, but that they can easily be trained not to do so. Once again, this shows that eyes do not automatically attract one’s attention.
Exploitation Route The broader impact of this work lies at the interface between human and machine vision systems. In fact, this is a topic which has received a lot more attention by engineers than psychologists. There are clearly many applications which benefit from being able to establish whether a person is present or not. To this end, a great many automatic system already exist, independently of any knowledge of the human case. This project has already established that some highly efficient techniques used by engineers, are not used by the human visual system. This knowledge will lead to predictions about the use of some automatic solutions – i.e. the range and limits of their utility. We have already begun discussions with a group of engineers which we hope will lead to further research proposals aimed at optimizing automatic face detection.
Sectors Security and Diplomacy

 
Description They haven't, really - except by other academics developing this field.
First Year Of Impact 2009
Sector Education
Impact Types Societal