Category: CMUTR
Institution: Department of Computer Science, Carnegie
        Mellon University
Abstract: One of the key tools in physics-based vision has been color 
        histogram analysis.
        But so far histograms have only been used for pixel grouping, 
        color analysis, and material type labeling.
        In this paper we present a new, quantitative model of histograms
        that yields a more complete description of scene properties.
        
        Color histograms were first used for image segmentation by 
        grouping similar-colored pixels.
        In the mid-1980s it was recognized that the color variation for
        inhomogeneous surfaces may be modeled as a regular physical 
        process with a planar distribution in color space.
        The identification of this plane and the vectors that define it
        leads directly to an analysis of object color and illumination 
        color.
        
        However there is much more to be said about color histograms.
        The colors do not fall randomly in a plane, but form clusters
        at specific points in color space.
        The location, dimensions, and orientation of these clusters a
        description of surface roughness and imaging geometry, as well 
        as an improved estimate of illumination color.
        
        Furthermore this type of analysis is not limited to simple
        images without interreflection.
        We show that an understanding of the histogram may be extended
        for those cases where interreflection is present, and that 
        additional, useful information may be obtained that is not
        available in simpler scenes.
Number: CMU-CS-91-203
Bibtype: TechReport
Month: nov
Author: Carol L. Novak
        Steven A. Shafer
Title: Anatomy of a Histogram
Year: 1991
Address: Pittsburgh, PA
Super: @CMUTR