This paper proposes a new, efficient, figure from ground discrimination met
hod. This algorithm is based on the assumption that background data feature
s can be more easily detected than figure data features, thus emphasizing t
he background detection task land implying the name of the method). Along t
he iterative labeling process, data features are sequentially and permanent
ly labeled as "background," while more global information is being collecte
d to assist with the coming decisions, until the process converges. This pr
ocedure creates a bootstrap mechanism which improves performance in very cl
uttered scenes. The method can be applied to many perceptual grouping cues,
and an application to smoothness-based classification of edge points is gi
ven. A fast implementation using a kd-tree allows one to work on large, rea
listic images. (C) 1999 Academic Press.