Speaker: David Forsyth
 
Affiliation: UIUC (CS)

Title:

Shading and Recognition – or – The First Mrs. Rochester

 
Abstract:
The shading on curved surfaces is undeniably a cue to shape.  Current computer
vision methods for analyzing shading use physically unrealistic models, have serious
mathematical problems, cannot exploit rich initial conditions  and are not reliable in
practice.  We introduce a novel method of accounting for variations in irradiance resulting 
from interreflections, complex sources and the like.  Our approach uses a spatially varying 
source model with a local shading model.  Fast spatial variation in the source is penalised,
consistent with the rendering community's insight that interreflections are spatially 
slow.  This yields a physically plausible shading model.  For inference, we  use a variational 
formulation, with a selection of regularization terms which guarantee that a 
solution exists. Our method is evaluated on physically accurate renderings of virtual
objects, and on images of real scenes, for a variety of different kinds of boundary 
condition.  Reconstructions for single sources compare well with photometric stereo 
reconstructions.
 
 
 
back