April 2002

BACK   

Variations in lighting can have a significant impact on the appearance of an object. In this talk I will discuss a novel characterization of this variability for the case of Lambertian (non-shiny) objects. First, we represent lighting using spherical harmonics, and describe the effects of Lambertian materials as the analog of a convolution; this is similar to working in the frequency domain in signal processing. We are then able to show that almost all the appearance of Lambertian objects is determined by the first nine components of the lighting when represented as spherical harmonics. This allows us to represent the set of images an object produces with an analytically constructed 9D linear space, explaining prior empirical results.

We can use these results to build efficient object recognition systems. Specifically, I will describe a face recognition system that uses 3D models to identify someone in a single image. Our method allows the system to handle a range of lighting conditions. I will also show how our results can be used to build 3D reconstructions of objects using multiple images taken with a fixed pose, but variable lighting. And I will briefly describe other recognition work on a web search engine for 3D computer graphics models, and on recognition without the use of 3D models.

BACK

Last Updated: April 17, 2002