Salient keypoint selection for object representation

Abstract

In this paper, we propose a keypoint selection scheme for SIFT and KAZE features and demonstrate their effectiveness in object characterization. The selection criterion rely on the detectability, distinctiveness and repeatability of the keypoints. These scores are combined to give a keypoint saliency score. The keypoints are ranked according to their saliency values and weak/irrelevant keypoints are filtered out based on a threshold value. These keypoints are further augmented with the keypoints obtained by applying SIFT to the texture map constructed using Gabor filter. The keypoint set represents the boundaries and object regions effectively. Experimental results validate the claims that the salient keypoints chosen by the proposed methodology are well suited for object representation.