Object Search via Random Context

Abstract

Accurately object searching plays an important role in computer vision. Retrieving and locating target objects in images are object searching's two sub-tasks. Aiming to promote the precision and recall of object searching, selecting appropriate image representation methods is the core issues. The representation method needs to provide enough discriminative features. Our approach adopts locality sensitive hashing method to extract enough sift features. The extracted features contain inliers and outliers. In order to distinguish them, random context confidence scores of features are computed. Our algorithm offers 3 benefits:1) A novel partition method is adopted to divide images. It is easy to be parallelized during computing contexts.2) A novel random points selecting method is adopted to avoids ill-defined boundary for target objects; 3) Multiple target objects in one image can be located by clustering all the features of each image with their coordinates. The experiment on a challenging Belgalogo dataset highlights the performance of our approach.