| |
Abstract:
The goal here was to measure the smallest discriminable face
change. A morphing algorithm mixed two faces in variable
proportions to create a series of synthetic faces that each
differed by a tiny amount. For each trial, a sample face was shown,
followed by two test faces presented sequentially, one of which
matched the sample and the other which differed slightly. The
subject indicated whether the first or second test matched the
sample. If the subject was incorrect, in the next presentation the
face difference was incremented, and if a correct choice was made
three times in a row, the difference was decremented.
Discrimination thresholds were calculated as the average of the
staircase reversal points. Staircases for four faces were randomly
interleaved. Sample duration was 1000 msec, test durations varied
from 1000-50 msec, and all stimuli were masked. To quantify
thresholds, each face was expressed as a vector of coefficients of
a principal component (eigenface) decomposition, and the vector
angle between a face and its just noticeable difference calculated.
The threshold face-difference angle was typically around one deg
for 1000 msec presentations and this threshold remained constant
down to about 100 msec, at which point it rose rapidly. The good
performance at short (100 msec) presentations indicates that a
serial feature-by-feature comparison, whether covert or by scanning
eye movements, isn't required for fine face discrimination.
|
