Bio: Jon Barron is 4th year PhD candidate at UC Berkeley, supervised by
Jitendra Malik. He is currently a visiting student with MIT’s vision
group. His research concerns intrinsic images, shape reconstruction,
and biomedical imaging.

Abstract: Traditional methods for recovering scene properties such as shape,
albedo, or illumination rely on multiple observations of the same
scene to over-constrain the problem (structure from motion,
photometric stereo, etc). Recovering these scene qualities given just
a single image seems almost impossible in comparison — the space of
albedos, shapes, and illuminations that exactly reproduce a single
image is vast. However, certain worlds are clearly more likely than
others: shapes tend to be smooth, albedos tend to be uniform, and
illumination tends to be natural.

We therefore pose this problem as one of statistical inference, and
define an optimization problem that searches for the *most likely*
explanation of a single image. To this end we present priors on shape
and albedo inspired by the models applied to natural images for
denoising or deblurring. Our resulting technique can be viewed as a
superset of several classic computer vision problems such as
shape-from-shading, “intrinsic images”, illumination estimation, and
color constancy. Our one unified technique appears to outperform all
previous algorithms for solving these constituent tasks, given only a
single image.