Face Identification Systems

1
Face Identification Systems
2
Face recall systems provide a visualisation of
the witness memory for a face. Include
Identikit, Photofit, sketching, computer systems
(e.g. Identikit 2000, FACES 3.0, E-Fit,
Evo-Fit). Problem usually poor at producing
recognisable likenesses.
"The man was a pretty odd-looking character and
we didn't get a good look at his face, but he
didn't look that odd," Mrs Rule said. "The man
in the picture has half an ear - he didn't have
half an ear. And his moustache wasn't like that.
"I don't think I've ever seen anyone who looks
like that in Stalham or anywhere else in my
life." She added "Apparently the problem with
the moustache was that the police only had long
moustaches on their computer so they had to sort
of chop it off at the ends." A spokesman for
Norfolk Police said "The E-fit image is
compiled as a result of the witness to the crime
giving as accurate and detailed description as
possible and how much they are able to recall."
3
The top ten world's worst photofits
(Mirror.ac.uk, 26/11/09)
4
Ellis, Shepherd and Davies (1975)
Target faces
Photofits constructed immediately after viewing
targets (Independent raters judged left group
to be better likenesses of the targets than right
group).
5
Ellis, Shepherd and Davies (1975) First
experiment Subjects saw two Photofit
faces. Reconstructed them using Photofit - very
difficult, even when the target Photofit face was
present. Second experiment Subjects saw 6
faces produced Photofits of them from memory,
immediately afterwards. A second group tried to
use these Photofits to identify the 6 original
faces from a set of 36 faces. Chance matching
performance 3 (1 in 36). Actual success rate
was 12.5 - better than chance, but very poor.
6
Ellis, Shepherd and Davies (1978) First
experiment Participants saw video of a man
reading (either 15 seconds or 2.5 minutes
long). Attended to the man's face (expecting to
make a Photofit of it) OR to the passage
(expecting a comprehension test). No benefits
from attending to the face. All Photofits were
independently rated as being poor likenesses.
7
Ellis, Shepherd and Davies (1978) Second
experiment Compared Photofit to freehand
sketching. Sketches were better than Photofits
if the target face was physically
present. Photofit was better when the target was
reconstructed from memory. Sketches were
"child-like in their simplicity and bore only a
vague approximation to the original target faces
- Photofits rated little better! Photofit was
still a poor likeness even when the face was in
view while it was constructed i.e., an accurate
COPY of a face is extremely difficult to produce
using Photofit.
8
Laughery and Fowler (1980) Compared sketching
with Identikits. Sketches were rated better
overall likenesses than Identikits. Sketches
showed a difference in rated likeness between
"from memory" and target face present"
conditions. No difference with Identikit -
because the likenesses were so poor.
9
Verbal descriptions (Christie and Ellis
1981). Better than Photofits. E-fit (Davies,
van der Willik and Morrison 2000). No better than
Photofit when face is constructed from memory -
can only produce good likenesses when the face is
in view at time of construction.
10
E-Fits made by trained police operatives with the
target face present throughout construction
11
Frowd et al (2005, 2007) forensically-relevant
evaluations 2005 compared Photofit, E-Fit,
Pro-Fit, Evo-Fit. Witness makes composite of a
face that is famous but unknown to them. Police
procedures followed, plus realistic delay between
seeing face and making composite. Different
participant tries to identify the famous face
(known to them). Various measures - spontaneous
naming most relevant. E-Fit and Pro-Fit best, but
only 20 of composites were recognised. 2007
compared E-Fit, Identikit 2000, FACES 3.0. Only
two out of 480 composites could be spontaneously
identified.
12
Why is FRS performance so poor? 1. Limitations
in eyewitnesses themselves. 2. Equipment
limitations (restricted feature-sets). 3.
Interference effects from Photofit features. 4.
Problems in method of construction. (verbal
mediation is difficult due to impoverishment of
descriptors, and may be prone to verbal
overshadowing). 5. Inappropriate
"feature-based" theoretical basis (Penry 1971),
at odds with the configural-based processing
actually used for face recognition.
13
Cheque cards with photographs (Kemp, Towell and
Pike 1999). Cashiers failed to detect mismatches
between user and photo on card. Falsely accepted
over 50 of fraudulent cards, and falsely
rejected over 10 of legitimate ones.
Mr. Cruise is here to see you
14
Recognition in surveillance videos (Burton,
Wilson, Cowan and Bruce 1999). Subjects judged
whether faces seen in high-quality photos had
been seen before in video clips. Subjects
personally familiar with the targets performed
well subjects unfamiliar with them performed
poorly.
15
Henderson, Bruce and Burton (2001)
Expt 1 Robber 1 Robber 2
First choice
correct 21 19
Incorrect "not present" 21 37
Picked someone else 58 43
Experiment 1 CCTV images
Expt 2 Robber 1 Robber 2
First choice
correct 16 40
Incorrect "not present" 44 51
Picked someone else 40 9
16
Interference effects from viewing other
faces/features Wogalter (1991) Participants
described a face by either (a) spontaneously
supplying applicable adjectives, or (b) ticking
applicable adjectives in a list of
alternatives. Group (b) were poorer at
recognising the face than (a). Perhaps
distracting adjectives contaminated face memory.
17
Problems with the method of construction Verbal
description and Photofit construction are
sequential, feature-by-feature processes, unlike
face representations. Hall (1976) Making verbal
descriptions of a face to a sketch artist
impaired subsequent recognition of the described
face. Christie and Ellis (1981) No correlation
between rated quality of a subject's Photofit
constructions and the quality of their verbal
descriptions.
18
Verbal overshadowing Verbally describing a face
may impair subsequent recognition of that face -
and others (Schooler and Engstler-Schooler 1990
Meissner and Brigham 2001). Elaborative
description produces more overshadowing (Meissner
and Brigham 2001). Effects cross semantic
categories e.g. describing a face impairs car
recognition (but not vice versa) (Brown and
Lloyd-Jones 2003). Perhaps verbal description
encourages inappropriate processing strategies
(e.g. local rather than holistic), which hinder
retrieval of face-appropriate information.
19
Verbal overshadowing and processing
orientation Macrae and Lewis (2002) Participan
ts tried to recognise a "robber" from an
8-alternative photo lineup. Interpolated task -
either (a) identify large letters comprised of
small letters ("global" orientation) (b)
identify the small letters, ignoring large
letters ("local" orientation). (c) control task,
reading a passage. (a) gt (c) gt (b).
Overshadowing is produced by an inappropriate
processing orientation - not "verbal" as such.
20
Evidence that face recognition involves more than
getting the right features in roughly the right
places
21
The composite face effect (Young, Hellawell and
Hay 1987). Upright faces are processed in an
integrated "holistic" way, that prevents easy
access to their constituent features.Inversion
abolishes this effect.
22
"Whole over part" advantage Features are
recognised better if they are presented within a
whole face than if presented in isolation or
within a scrambled face (Tanaka and Farah 1993).
23
The Bruce and Young (1986) model of face
processing
Structural encoding
Recognition
Expression Facial Speech Age Gender
Face Recognition Units
Person Identity Nodes
Name Generation
24
Stages in Face recognition
Structural encoding Based on features, and their
configuration (spatial relationship)
Face Recognition Unit Activated by a match to a
stored face representation
Person Identity Node Contains semantic
information about the person
Name Generation
25
Familiar vs unfamiliar face recognition Both
familiar and unfamiliar face recognition involve
simliar types of processing (Collishaw and Hole
2000). But - familiar faces recognised better
from internal features, unfamiliar faces from
external features. (Ellis, Shepherd and Davies
1979 Young, Hay, McWeeny, Flude and Ellis
1985). Familiar face recognition based on
"abstractive" representation compiled from many
views unfamiliar face recognition is more
image-based/episodic (Burton et al 2005 Megreya
Burton 2006, 2008).
26
Conclusions 1. Face processing involves
configural processing face recall systems need
to be sympathetic to this. 2. Future systems
could take into account the biological
constraints on what can occur in a face - e.g.
dolicocephalic versus brachycephalic (Enlow 1982).
27
3. Using multiple composites might aid
recognition (a) Brace et al (2006) Presenting
witnesses with any four composites (from 8) aided
recognition.
28
(b) Bruce et al (2002) Morphs of 4 witnesses'
composites, and sets of four composites,
recognised better than individual composites.
Condition Correct False positives
4 composites 38 0
Morph of 4 composites 28 6
Best single composite 22 0
Worst single composite 16 6
29

• 4. Using multiple techniques can improve
  recognition
• Frowd, Skelton, Hassan, McIntyre and Hancock
  (2013)
• EvoFIT alone 8-20 recognition rate after 3-4
  hours delay, even worse after 3 days.
• EvoFIT plus
• Cognitive Interview of witness, focusing on
  holistic aspects of face (attractiveness,
  kindness, etc).
• Blurring of external features
• (iii) Vertical stretching of finished composites
• Mean recognition rate increased to 74