Graph Cut Segmentation for Improvement of the Selective Search in Object Detection

Abstract

Object detection usually requires determining presence of a specific object for a huge number of windows, which is exhaustive but inefficient. To improve object detection system, researches on how to effectively propose a small number of object location hypotheses to an object detector are intensively conducted for recent several years, which aims to speed up object recognition and replace the existing sliding window method. Similarity grouping method based on hierarchical grouping in object proposals has limits that it locally merges neighboring regions, so it shows low performance when objects contain various color, texture or appearance of an object is similar to that of background. To solve the problem, we perform a graph cut segmentation in the middle of hierarchical grouping, and extract additional object location hypotheses for foreground regions obtained from graph cut segmentation. Graph cut segmentation can consider whole regions of an image by minimizing energy function. Experiments with Pascal VOC 2012 dataset show that our proposed method shows improved performance of proposing object hypotheses to a detector.