Conditional Random Field based salient proposal set generation and its application in content aware seam carving

Abstract

Most of the existing generic object localization algorithms usually give the plausible object locations without taking into consideration the saliency ordering of the proposal set. This paper presents a novel object proposal generation which ranks the key objects according to their saliency score in the proposal pool. First, we formulate a Bayesian framework for generating a probabilistic edgemap which is used to assign a saliency value to the edgelets. A conditional random field is then learnt for edge-labeling by effectively combining the edge features with the relative spatial layout of the edge segments. Lastly, we propose an objectness score for the generated proposal set by analyzing the salient object edge density completely lying within the candidate boxes. Extensive experiments on the benchmark PASCAL VOC 2007 and 2012 datasets demonstrate that the proposed method provides competitive performance against popular generic object detection techniques while using fewer number of proposals. Additionally, we demonstrate the applicability of the generated proposal set for content aware image retargeting.