110 likes | 246 Views
What do color changes reveal about an outdoor scene?. Kalyan Sunkavalli Fabiano Romeiro Wojciech Matusik Todd Zickler Hanspeter Pfister Harvard University. 報告者:黃智勇. CVPR’08. 指導教授:林惠勇. Outline. Introduction A color model for outdoor image sequences Incorporating shading Model fitting
E N D
What do color changes reveal about an outdoor scene? Kalyan Sunkavalli Fabiano Romeiro Wojciech Matusik Todd Zickler Hanspeter Pfister Harvard University 報告者:黃智勇 CVPR’08 指導教授:林惠勇
Outline • Introduction • A color model for outdoor image sequences • Incorporating shading • Model fitting • Experimental Results • Discussion
A color model for outdoor image sequences • Restricting our attentionto Lambertian surfaces and linear sensors, the trichromaticobservation of any surface point under illuminantE(λ) can be written as • We choose linear transforms as our modelfor the effects of illuminant changes, we assume that the subspace containingdaylight spectra is two-dimensional. According to this assumption,the observation of any given material under anydaylight spectral density can be written as [7, 4]
Incorporating shading • Assumethat the scene is static, that reflectance at scene pointsis Lambertian, and that the irradiance incident at any scenepoint is entirely due to light from the sky and the sun.
Model fitting • When the scene contains foreground objects, interreflections,and non-Lambertian surfaces, estimates of the colorplane for each pixel (i.e., the normals u3(x)) can be corruptedby outliers. In these cases, we have found that enforcingEq. 6 as hard constraints yields poor results. A betterapproach is to perform an unconstrained minimizationof the objective function in Eq. 7
Color Constancy our use of general linear transforms can be expected to provide increased accuracy over what could be obtained using common diagonal or generalized diagonal transforms [4].
Background subtraction Shadow Prediction
Discussion • Canbe used for a variety of visual tasks. Examples in this paperinclude color constancy, background subtraction, andscenereconstruction . • The model could be improved by incorporating robustestimators into the fitting process, by using a more flexiblereflectance model, and by making use of temporal patternsto appropriately handle ’time-varying textures’ suchas moving water and swaying trees.