| |
Abstract:
We develop a hierarchical generative model to study cue
combination. The model consists of four layers: global shape
parameters at the top, followed by global cue-specific shape
parameters, then local cue-specific parameters, ending with an
intensity image at the bottom. Inferring parameters from images is
achieved by inverting this model. Inference produces a probability
distribution at each level; using distributions rather than a
single value of underlying variables at each stage preserves
information about the validity of each cue for the given image,
which helps make this model more powerful than standard linear
combination models or other existing combination models. The
parameters of the model are determined using data obtained from
psychophysical experiments. In these experiments, subjects estimate
surface shape from intensity images containing texture information,
shading information, or both types of cues, and varying degrees of
noise. This model provides a good fit to our data on the
combination of these two cues, and also provides a natural account
for many aspects of cue combination.
|
