Transcript Reflectance Function Approximation

Reflectance Function
Approximation
Ted Wild
CS 766
Thursday, December 11, 2003
Motivation
• Material recognition
• Classification Problem
– Dror et al.
– Recognizing materials with known reflectance
functions
– CUReT, Dana et al.
Example
• Denim + Cotton +
Skin = Person
• Can make feature
tracking,
segmentation, etc.
easier
• Would work under
different pose, lighting
Reflectance
• How light and surface
interact
• Depends on angle at
which light hits
surface and view angle
• Figure by Wallace and
Price
Bidirectional Reflectance
Distribution Functions
• Scalar function of 4 variables:
– Incident light (2 angles)
– View direction (2 angles)
• CUReT data
– BRDF’s of 61 materials
– 205 measurements per material
CUReT
CUReT
CUReT
BRDF Approximation
• Kernel regression
• Gaussian kernels
• 2 parameters
•
min  K(A, A')  be  y  
(K(A,B)) ij  e
  A i 'B j
2
Approximation Results
BRDF Classification
• Given:
– Set of known BRDF functions
– Set of BRDF measurements for a material
• Determine what material the measurements
are taken from
Method
• Approximate known reflectance functions
from data
– Kernel regression
• Use nearest-neighbor classification to
identify new function
• Evaluation: Leave out random data from
CUReT measurements, try to classify left
out data
Questions
• How accurate does reflectance function
approximation have to be for classification?
• How many points are needed to get
sufficient accuracy?
– Known BRDF approximation
– Classification
• What models of reflectance work well?
Classification Results
Problems
• Need to know:
– Geometry
• Discussed in class
– Illumination
• Ramamoorthi and Hanrahan
• Factorization technique to recover BRDF and
lighting in some cases
• Can only recognize “known” materials
Recognizing Unseen Materials
• If input is sufficiently different from known
BRDF’s, create a new class for it
• Use linear combination of known BRDF’s
for further recognition
– May need less points for recognition than for
approximation
– Can improve approximation of new class as
more of its measurements are identified
Very Early Results
• Leave one material out:
– When testing, classify material as unseen if the
distance to its nearest neighbor >= tol
• tol = 0.25
– Average error: 0.40, Predicting unseen: 0.51
• tol = 0.20
– Average error: 0.45, Predicting unseen: 0.29
• Trials only run once!
Influences
• Dror et al. (2001)
– Material classification based on reflectance
• Lensch et al. (2001)
– Representation of BRDF’s as combination of a
few basis BRDF’s.
• Dana et al. (1997)
– Use of CUReT data to evaluate reflectance
function approximation
Future Work
• Complete and test method for unseen
material recognition
• Reduce error for approximation and
classification methods
• Recognition of materials under unknown
geometry and/or illumination
• Evaluate other reflectance models