15-869 Project Progress Report
Title: Illuminating the Face
Terence Sim
Nov 15, 1999

Problem:

Given 1 face image taken under single point light source (unknown),
render same face under new lighting direction.
 
(Col. 1) Input image
(Col. 2-4) face under new lighting directions

Related Work:

Quotient Image: A.  Shashua (CVPR99)
  • illumination-invariant image
  • uses 1 image


    • Illumination cone: P. Belhumeur (CVPR98)

  • recover face shape
  • uses 3 images

    • Method:

    What's Novel:

  • Probabilistic approach
  • E(s) term models shadows/specularities, learned from training set


    • cf:
    Shashua:

  • assumes all faces have same surface normals
  • cannot incorporate > 1 input image
  • sensitive to training set
  • cannot handle shadows/specularities


    • Belhumeur:

  • requires at least 3 input images

    • Results:

    Input image:

    Synthesized images:
    Input image as light source moves from left to right,top to bottom.
    The input image is taken under (0,0)