Are two views better than one? Investigating three-quarter view facial composites
| Pages | 291-306 |
| Published date | 09 November 2015 |
| DOI | https://doi.org/10.1108/JFP-10-2014-0040 |
| Date | 09 November 2015 |
| Author | Hayley Ness,Peter J.B. Hancock,Leslie Bowie,Vicki Bruce,Graham Pike |
| Subject Matter | Health & social care,Criminology & forensic psychology,Forensic practice |
Are two views better than one?
Investigating three-quarter view
facial composites
Hayley Ness, Peter J.B. Hancock, Leslie Bowie, Vicki Bruce and Graham Pike
Dr Hayley Ness is Lecturer in
Applied Cognition at the
Open University in Scotland,
Edinburgh, UK.
Professor Peter J.B. Hancock
is based at the Department of
Psychology, University of
Stirling, Stirling, UK.
Dr Leslie Bowie is Chief
Technology Officer at ABM (UK)
Ltd, Beeston, Nottingham, UK.
Professor Vicki Bruce is based
at the School of Psychology,
Newcastle University,
Newcastle Upon Tyne, UK.
Professor Graham Pike is
based at the Department of
Psychology, Open University,
Milton Keynes, UK.
Abstract
Purpose –The introduction of a three-quarter-view database in the PRO-fit facial-composite system has
enabled an investigation into the effects of image view in face construction. The purpose of this paper is to
examine the impact of constructing full-face and three-quarter-view composites under different encoding
conditions. It also examines the potential value of three-quarter-view composites that can be generated
automatically from a front-view composite. The authors also investigate whether there is an identification
benefit for presenting full-face and three-quarter composites together.
Design/methodology/approach –Three experiments examine the impact of encoding conditions on
composite construction and presentation of composites at the evaluation stage.
Findings –The work revealed that while standard full-face composites perform well when all views of the
face have been encoded, care should be taken when a person has only seen one view. When a witnesshas
seen a side view of a suspect, a three-quarter-view composite should be constructed. Also, it would
be beneficial for a witness to construct two composites of a suspect, one in full-face view and one in
a three-quarter-view, particularly when the witness has only encoded one view.No benefit emerged for use of
three-quarter-view composites generated automatically.
Research limitations/implications –This is the first study to examine viewpoint in facial composite
construction. While a great deal of research has examined viewpoint dependency in face recognition tasks,
composite construction is a reconstruction task involving both recall and recognition. The results indicate that
there is a viewpoint effect that is similar to that describedin the recognition literature. However, more research
is needed in this area.
Practical implications –The practical implications of this research are that it is extremely important for facial
composite operators in the field (police operators) to know who will make a good likeness of the target.
Research such as this which examines real-life issues is incredibly important. This research shows that if a
witness has seen all views of a perpetrator’s face then standard composite construction using a full-face view
will work well. However, if they have only seen a single view then it will not.
Social implications –There are obvious wider societal implications for any research which deals with
eyewitness memory and the potential identification of perpetrators.
Originality/value –No research to date has formally examined the impact of viewpoint in facial-composite
construction.
Keywords Viewpoint, Eyewitness memory, Facial composites, Forensic cognition, PRO-fit, Three-quarter-view
Paper type Research paper
Introduction
Computerised composite systems assist witnesses in constructing a facial likeness of a suspect.
In the UK, E-FIT and PRO-fit are used as well as newer, more sophisticated systems such as
EvoFIT (Frowd et al., 2013) and EFIT-V (George et al., 2008).
Received 24 October 2014
Revised 4 April 2015
28 April 2015
9 September 2015
Accepted 9 September 2015
The authors would like to thank
Krista Overliet and Sivakumar
Anandaciva who assisted with the
data collection for experiments
one, two and four. The authors
would also like to thank Michael
Bromby, formerly of ABM UK for
his technical assistance with
PRO-fit and Duncan Robb and
Jane Turnbull at Queen Margaret
University College, Edinburgh for
their assistance with the
recruitment of participants.
DOI 10.1108/JFP-10-2014-0040 VOL. 17 NO. 4 2015, pp. 291-306, © Emerald Group Publishing Limited, ISSN 2050-8794
j
JOURNAL OF FORENSIC PRACTICE
j
PAG E 29 1
Research has demonstrated that computerised systems produce better likenesses than older
systems such as Photofit (e.g. Brace et al., 2006a; Cutler et al., 1988; Davies et al., 2000; Koehn
and Fisher, 1997; Wogalter and Marwitz, 1991) and newer systems can produce very good
likenesses under some circumstances (e.g. Frowd et al., 2005a, b, 2013). However, despite
these improvements, composites can still portray a poor resemblance to the target. This is
because constructing a composite is a difficult task. Witnesses remember the face of someone
they have only seen once, perhaps for a very short time. The resulting image is circulated such
that someone who is familiar with the suspect will recognise him or her from the composite. Due
to their forensic importance, it is vital then that researchers examine methods for improving the
identification of composites.
Some of this research has improved the effectiveness of composites after they have been
constructed. For example, researchers have examined morphing composites from single and
multiple witnesses (e.g. Brace et al., 2006b; Bruce et al., 2002), caricaturing composites (Frowd
et al., 2007a) and by “splitting”a composite horizontally (McIntyre et al., in press). Although this
research has had success in improving recognition rates, no research has examined the impact
of viewpoint on both construction and subsequent identification of a composite image. In this
context, viewpoint refers to the relative horizontal rotation of the target face to the viewer (e.g. full-face,
three-quarter and profile views).
Consideration of viewpoint is likely to be important. In a real-life situation, witnesses will have seen
a face that is usually unfamiliar, three-dimensional (3D) and in motion; however, when
constructing a face, the result is a two-dimensional static representation. Evidence on the role of
movement for unfamiliar face recognition has indicated that non-rigid (natural) motion at encoding
may help us to form a robust 3D representation (e.g. Bruce and Valentine, 1988; Pike et al., 1997;
Schiff et al., 1986). As witnesses may have encoded such a representation of the face, what is the
best viewpoint to use? Would it be easier for them to construct a three-quarter-view composite
rather than a full-face composite?
Indeed, witnesses may have only seen a side view of the face, and so is it appropriate to ask them
to construct a composite in a different view? Frowd et al. (2014) indicate not, as such changes
reduce composite identification (for a holistic system). Also, one might question whether
constructing a composite in a single view sufficiently captures the three-dimensional nature of the
face to facilitate later identification of the composite image? While research has not examined
both issues for composite construction, researchers have investigated whether one particular
view is preferred in face recognition (e.g. Bruce et al., 1987; Hill et al., 1997; Lui and Chaudhuri,
2002; Newell et al., 1999; Schyns and Bülthoff, 1994). This research stems partly from research
on object recognition that has suggested not only that object recognition may be viewpoint
dependent (e.g. Edelman and Bülthoff, 1992; Tarr and Pinker, 1990), but also that certain views
of an object are preferred (e.g. Palmer et al., 1981).
For recognition of faces, as a three-quarter-view is centred between full-face and profile, it may
contain informati on that is available in b oth views. As such, a thr ee-quarter-view c ould be a
canonical representation. When multiple views of a face are seen at study, some researchers
report that profile views (90°) perform poorly but recognition performance for full-face (0°) and
three-quarter-views (45°) do not differ significantly(Hill et al., 1997; Newell et al., 1999). Similar
results have been obtained by Logie et al. (1987) using “live”targets. Bruce et al. (1987) also
found a three-quarter-view advantage for unfamiliar but not familiar faces; Baddeley and
Woodhead (1983) and Krouse (1981) report a similar three-quarter-view advantage for
unfamiliar faces.
This suggests that there may be a three-quarter-view advantage when more than one view is
presented at study. However, Liu and Chaudhuri (2002) and others (e.g. Davies et al., 1978;
Laughery et al., 1971) have failed to find an advantage. Indeed, Schyns and Bülthoff (1994)
compared two different side views (18°, 36°) with a full-face view and found that no one view was
preferred. It seems that when multiple views are encoded at study, recognition performance is
equivalent for both full-face and three-quarter-views.
When single full-face images are seenat study, researchers have failedto find a three-quarter-view
advantage at test (Newell et al., 1999; Patterson and Baddeley, 1977; Woodhead et al.,1979).
PAGE292
j
JOURNAL OF FORENSIC PRACTICE
j
VOL. 17 NO. 4 2015
To continue reading
Request your trial