Occlusion boundary detection and Figure/ground assignment based on correlation clustering algorithm in a single image

Abstract

Occlusion boundary detection and figure/ground assignment are essential step for understanding a scene through recognizing objects in single image. A single image contains 3D objects which are projected on 2D plane. Consequently, objects are occluded each other losing their own information. This makes object recognition and scene analysis dicult. However, by knowing that occluded boundaries and their figure/ground relationship we can infer the 3D structural relationship between the objectsand distinguish the objects with others. Therefore, we propose an algorithm to assign figure/ground labels on occlusion boundary by utilizing task-specific image partitioning method(TSP) and correlation clustering algorithm. Our framework is a two-step approach : First, we generate segmented image by using TSP based on higher-order task-specific superpixels image(HOTSS). Second given segmented image which conserves occlusion boundary, a modified correlation clustering algorithm takes a decision to each occlusion boundary whether front or back. We solve this as a energy minimization problem by using alinear discriminant function which allows for polynomial-time inference by linear programming (LP) and large margin training based on structured support vector machine(S-SVM). We evaluate the proposed algorithm on the Gemetric Context Datasets of various outdoor images taken from common view-point using Google image search. The experiments show state of the art results both using ground-truth segmentations (93.0%) and task-specific segmentations (80.15%).