Title: A fast self-organized iconic segmentation and grouping based in color 
Javier Finat, Margarita Gonzalo-Tasis and Alejandro Viloria. 
Javier Finat, Dept of Algebra and Geometry, ETS Computer Science Eng, Univ of Valladolid, 47011 Valladolid, 
Spain, e-mail: jfinat@agt.uva.es 
Keywords: Vision Sciences, Representation, Interfaces, Color, Visualization, Motion 
Abstract 
Color is a powerful tool for segmentation, recognition and tracking of mobile objects. In this work we address the clustering problem 
to the iconic level for the static and dynamic case. After thresholding, averaged vector sums simplify the grouping of regions around 
typical color values. Our iterative algorithm follows a spatial propagation scheme inspired on a WTA variant of a Self-Organizing 
Map; so, we obtain a segmentation linear in the number of regions. The user can fix this number in advance too. The linear character 
allows us to apply this approach to obtain a coarse depth map and a mobile segmentation at iconic level without extracting geometric 
characteristics of boundaries. Nevertheless, our procedure compares merged regions located at successive images of a movie-file 
taken near in time. This comparison 1) gives an optimal simplicial Delaunay decomposition in the color space, 2) evaluates some 
characteristics of the movement, 3) discriminates between egomotion and movement of some object in the scene, and 4) captures 
iluminance variations as color variation 
1. Introduction 
Color phenomena arise from an interaction of light with matter. 
Low-level vision based in color gives us a local 
homogeneization for a vector representation of RGB or IHS- 
color associated to the matrix of intensity functions. 
Segmentation based in homogeneous color regions ([Oht80]) 
simplifies the tracking through sequences of images ([Fer92], 
[Hea92]). Geometric aspects linked to the relative orientation 
would must be as independent as possible of the illumination 
conditions, but they are sensitive to shadows and reflections; 
furthermore, they are blurred by global filters which degrade the 
quality and continuity of boundaries. Luminance and IHS-color 
properties give us a more realistic feeling, because Hue and 
Saturation are less sensitive to illumination variations; in fact, 
they are invariant to surface curvature and lighting conditions 
([Hun95]). 
Here, we present a bottom-up strategy to segmentation 
problems. Our goal is to avoid as much as possible any 
dependence w.r.. the real scene and the illuminationo. 
Traditionally, a bottom-up approach to the image segmentation 
includes the smoothing processes, chains of neighbour points, 
connectivity detection, correction of aliasing effects, matching 
the segment model in a database to generate the geometrical 
model of the scene and to complete hidden zones. The color 
segmentation uses color to separate objects without a priori 
knowledge about specific surfaces. Inversely, the color 
recognition attempts to recognize colors of known color 
characteristics. Thus, the color segmentation seems more 
appropriate for a bottom-up approach. 
There are different approaches to the color segmentation 
depending on geometrically- or physically-based methods. 
Geometric methods are classified after the selected dimension 
for grouping criteria OD or pixel level ([Che98]), 1D or edge 
level ([Wu93]), and 2D or region level ([Gau99]). Furthermore, 
there are hybrid approaches which are successful for complex 
images ([Shi99]). Our hybrid strategy intends to integrate all 
geometric levels of the analysis in a common data structure with 
optimal and easily updatable properties. To achieve this gaol, 
we have chosen a classical approach linked to the 
superimposition of an automatically generated geometric 
construction based in Delaunay decompositions as an optimal 
support for color configurations. Optimal properties of 
Delaunay decompositions are well-known ([Ede92]), but 
seemingly this viewpoint has not been considered before. 
2. Computational Background 
There are several algorithms to retrieve and cluster regions 
inside a image, but usual algorithms are of recursive nature, and 
they represent a high comptuational cost, often. A region can be 
defined as a set of adjacent pixels with similar color. This 
region is labelled as a unit, with a localization (position- 
orientation for its centroid), and an averaged region color in 
some vector representation. All pixels belonging to a particular 
region are labelled with the same value. There are no different 
regions characterized with the same label. 
The output of a low-level geometrical aproximation is given by 
a collection of segments which must be grouped to obtain 
boundaries and contours of the objects represented in the image. 
Inversely, if regions are retrieved easily, the contour model can 
be generated in a faster way. Often, it is not necessary to have a 
precise mathematical knowledge of contours appearing at 
images; it suffices an iconic information. Alternately, we can 
use a clustered region model directly, instead of using a contour 
model. In this case, the computational cost is lower, because we 
don’t need to have precise information about the mathematical 
primitives and their projections. We have chosen the second 
approach to avoid the computational cost of boundaries 
extraction in complex images and to use the discrimination 
power of color, as in the human case. 
Usual approaches are based on a two-step recursive (to identify 
the averaged color) and iterative (for regions labelling) 
—152—