|
I am a computer science
graduate student at Carnegie
Mellon
University
studying computational and biological vision. I am interested in the
statistics of natural scenes and the use of these statistics to help
infer scene properties from images. My current work is the inference of
3D shape from single images. Most existing approaches to shape
inference work by starting with theoretical, physics-based models of
image formation and then inverting these models. Unfortunately,
inverting the image formation process is highly underconstrained.
This forces us to revert to oversimplified models of image formation
which may be unrealistic in natural scenes. Various assumptions about
image formation parameters have to be made, such as Lambertian (matte)
surface reflectance, uniform albedo, single point illumination,
infinitely distant illumination, smooth 3D surface shape, the absence
of shadows, the absence of interfacet reflection, and others. However,
these assumptions are often violated in the real world, and this leads
to poor generalization for these algorithms.
I believe that, in addition to clarifying the relative merit of these
assumptions, a solid understanding of the statistics of natural scenes
will uncover new, exploitable regularities in image formation that are
not obvious from physical models. One of the earliest discoveries in
our database was a direct anticorrelation between distance and
brightness: darker image regions are more likely to be further away.
That brighter objects appear closer was first observed by da Vinci. Our
database provides the first evidence that this relationship holds in
natural scenes. We believe that this trend is attributable to shadows:
image regions that lie within object interiors, crevises, or
concavities are farther from the observer than object exteriors, and
the object interiors are more likely to lie in shadow. Additional
exploitable statistical trends may result from regularities in the 3D
shape of objects, regularities in their spatial relationships,
regularities in the location and orientation of the observer,
regularities in illumination conditions, etc. Despite the potential
usefulness of statistical models, and the growing success of
statistical methods in vision, few studies have been made into the
statistical relationship between images and range images.
|