A region extension method for natural outdoor scene interpretation

Abstract

This paper describes a method for natural scene interpretation which first extracts core regions of high reliability using local feature default models and qualitative constraints and then carries out final scene interpretation by extending these core regions to true object boundaries. Final scene interpretation proceeds by processing core regions with highest reliability first. For the determination of true object boundaries, maximal bounds for the object regions are determined using a hue depletion method and statistically robust curve fitting, followed by a step of compensating for local feature shifts based on hue impletion. This method allows the interpretation of difficult scenes in spite of local feature variations and weak object class separation.