Spatial Context in an Image Analysis System

Abstract

“Scene Analysis”, especially for real data, is a complex problem. There are two main explanations which interest us in this paper. The first one is that sometimes an object identification needs informations about his spatial context ([GAR76,OHTA89]). We define the spatial context of an object as topological relations beetween this object and the other objects of the world. The second explanation is that the detection of an object implies to solve simultaneously two problems, the localization and the identification. This is really difficult because sometimes for the same object class, we must consider variations, for instance shape variation or colour variation. To solve these problems, we use generic models of objects [FUA87,GAR89] which can be expensive with computing time if they explore the whole scene. In this paper, we increase the formalization of spatial context and we show how it allows to focus the search objects in a limited aera of the scene.