|
This page presents a collection of experimental results, obtained by application
of the proposed
bottom-up discriminant saliency detector (DSD) to a set of displays,
classically used in psychophysical studies of human saliency and visual search.
Single and conjunctive feature search
The discriminant saliency maps are shown below each of the displays.
In displays (a) and (b), targets differ from background in terms of a single
basic feature (orienation (a) and color (b)), and pop-out. In display (c),
the target is defined by a conjunction of orientation and color, and does
not pop-out.
Search asymmetries and Weber's law
The following examples show that discriminant saliency predicts asymmetries
commonly found in visual search experiments: a target pops-out if it is defined
by the presence of a feature that is absent in the distractors (left display),
but the reverse does not hold (right display).
Furthermore, discriminant saliency predicts the compliance of search asymmetries
with Weber's law, as proposed by Treisman & Gormican (1988; Feature analysis in
early vision: evidence from search asymmetries. Psychological Review, 95,
14-58, 1988). In the example below, a display is shown on the left, where the
target and the distractor are both vertical segments which differ only in length.
On the right, we present a scatter plot of the value of discriminant saliency as a
function of the ratio DX/X
(where DX is the difference between
target and distractor length, and X is the distractor length), which appears
in Weber's law. The dashed line shows the best fit to Weber's law.
Nonlinearities of human saliency
In this experiment discriminant saliency is measured as a function of orientation
contrast (orientation difference between target and distractors). Three displays
of different orientation contrast are shown in the top row. The two plots in the
bottom row present the saliency measured for both human subjects
(left, Figure 9 of Nothdurft (1993); The conspicuousness of orientation and
motion contrast. Spatial Vision, 7, 341-363, 1993) and discriminant
saliency (right). Both plots show strong threshold and saturation effects,
i.e. a sigmoidal shape, showing that discriminant saliency makes accurate
quantitative predictions of human saliency perception.
Distractor heterogeneity and search surface
Nothdurft (1993) showed that when the distractors in the above displays become
heterogeneous, the percept of target saliency by human subjects is significantly
reduced. In particular, he measured human saliency for three levels of distractor
heterogeneity, using the displays shown in the top row of the following figure.
The human saliency curves are shown in the left plot of the bottom row.
On the right, we present the measure of discriminant saliency on these displays.
Once again, discriminant saliency has a very good match to the human data.
Furthermore, by plotting the discriminant saliency curves in 3-D (left plot below),
it is clear that discriminant saliency reproduces the "search surface" proposed by
Duncan and Humphreys (1989; Visual search and stimulus similarity.
Psychological Review, 96, 433-458, 1989), shown on the right.
|
