Datensatz

Zusammenfassung

When we look at everyday things, we not only perceive their 3D shape and identity, but also we generally enjoy a distinct visual impression of their material properties. Without touching
an object, we can usually tell at a glance whether it is soggy or dry, soft or hard, smooth or
rough. How does the visual system recognize materials? What are the major challenges that
the visual system faces? Despite the enormous variety and vividness of material perception, it
is only just beginning to emerge as a major topic of study in vision research [e.g. 1–7]. Here
I present a framework for understanding human visual perception of materials and introduce a
number of novel demonstrations of our visual aptitude for estimating material properties.
The main thesis I will present is that despite our exquisite sensitivity to changes in physical
states (e.g. subtle changes in appearance allow us to tell the difference between fresh and stale
bread) the visual system does not generally estimate the intrinsic physical attributes of materials
(e.g. density or coefficient of viscosity). Instead, it adopts a heuristic strategy for classifying
material appearance based on the statistical behaviour of materials in our environment. In
the real world, the observed appearance of a material is subject to constraints such as gravity
and natural illumination conditions. Consequently, there exists a large set of simple low-level
image measurements (e.g. contrast distributions, amplitude spectra, optic flow patterns, etc.)
that reliably correlate with changes in physical state.
I will present a taxonomy that organizes these cues, and embed all extant research on
material perception within this framework. Broadly I organize the cues into three classes:
(1) optical cues, i.e., information arising from the manner in which a material interacts with
light, such as its specular reflectance, or sub-surface scattering coefficients. This class of cues
has received the most attention, as there is a considerable body of work on the estimation of
diffuse albedo and colour. (2) geometric cues, i.e., the characteristic 3D shapes adopted by a
material subject to natural forces. (3) dynamic cues, i.e., the way that a material tends to change
shape over time or interact with other objects in the scene. I use a variety of physics-based
computer graphics simulations to demonstrate these cues and their low-level correlates. Finally
I show circumstances under which high-level (cognitive) factors can influence our perception
of materials.