CSAIL Event Calendar: Previous Series
Estimating Surface Reflectance Properties from Images
Speaker: Ron Dror , Stochastic Systems Group, Perceptual Sciences Group
How do you tell a sheet of white paper from a polished mirror? Humans can distinguish between the two materials at a single glance, but the difference in appearance is difficult to quantify. An image of either surface, especially the mirror, depends heavily on the pattern of light incident on the surface from various directions. An ability to identify reflectance properties visually would facilitate material recognition, geometric reconstruction, and motion estimation in computer vision systems. It would also provide a convenient method to capture surface properties essential to realistic computer graphics. We are developing techniques to determine the reflectance properties of surfaces from real images taken under uncontrolled illumination. We claim that humans succeed in this task only because real-world illumination possesses a great deal of predictable statistical structure. In order to take advantage of this structure, we learn relationships between surface reflectance and the wavelet statistics of appropriately warped surface images. We show preliminary results and discuss potential extensions to surfaces of more complex or unknown geometry.