| [back]
Perception
of Mirror Surfaces
Silvio Savarese, Fei Fei Li, Pietro Perona
Abstract.
The aim of our work is to investigate how the human visual system perceives
specular surfaces and which cues can be used to recover the shape of
such class of objects. Our experiments show that mirror reflections
are a weak cue for most human observers when additional information
is not available.
Introduction and Motivation. A sense of three-dimensional shape
may be perceived by looking at the photograph of the object. This perception
is due to different cues such as contour, shading, perspective and occlusion.
When looking at a photograph of a specular object, such as a silver
plate, a metal spoon or a car surface, one additional cue is represented
by the reflection of the environment around the object. A deformed picture
of the environment is seen on the surface of the object and the amount
and the type of deformation depend upon its shape. The final goal of
our research is to understand how the human visual system uses this
clue in the perception of shape.
Our Experiment. In order to investigate this question we asked
a number of human observers to discriminate between images of mirror
surfaces of qualitatively different shapes: a sphere, a cylinder and
the neck of a vase (Fig. 1). Such shapes have positive, zero and negative
Gaussian curvature and reflect the same scene with distinctly different
distortions. The experimental stimuli were 144 photographs of large
patches of each mirror surface reflecting one of six regular patterns
(Fig.2) which had been shown to the subjects in advance. Each patch
was obtained by vignetting one of the photographs using irregularly
shaped boundaries, in order to eliminate occluding boundary information
(Fig.3). It was viewed monocularly and centrally on a standard computer
monitor for either 1 or 5 seconds. Each patch subtended in average a
visual angle of 20 degrees. The subjects were instructed to respond
to three forced alternative choices (sphere, cylinder, vase) (fig.4).
Figure
1.
The 3 shapes used in the experiment.
Figure
2. The
6 patterns used in the experiment
Figure
3. (a).
The setup: a camera takes a picture of a mirror shape reflecting a pattern.
(b) Picture of the mirror shape reflecting the pattern: a patch of the
reflection of the pattern is cropped from th picture. (c) The cropped
patch does not include any visual cues besides a portion of the reflected
pattern.
Figure
4. Top:
Each trial of the experiment is carried out as follows: A fixation cross
is presented at the beginning of the trial for 240 msec. A stimulus
(see Bottom panel) is then presented for either 1 second or 5 seconds.
Subjects are instructed to respond the shape of the mirror as fast and
as accurately as possible by pressing one of the 3 designated keys.
Bottom (a,b,c) Examples of stimuli presented during experiment 1, 2,
and 3 respectively.
Results. Our subjects are only slightly better than chance in discriminating
such shape differences. Our ideal observer analysis indicates that mirror
reflections allow recovery of depth, tangent plane and surface curvature
when the surrounding world has a known shape. However, when the surrounding
world is (partially) unknown, the problem is underconstrained and many
solutions are possible. Our experiments confirm this analysis indicating
that mirror reflections are a weak cue for most human observers when
additional information is not available.
Figure
5.
This result indicates that our subjects are only slightly better than
chance in discriminating the three mirror shapes; 33% means chance level
performance.
S. Savarese, F.F. Li
and P. Perona, "Can We See the Shape of a Mirror", in Proc.
of 3rd Annual Meeting of Vision Sciences Society, 2003
top
|
