Incorporating local and global information using a novel distance function for scene recognition
In the field of scene recognition using only one type of visual feature is not powerful enough to discriminate scene categories. In this paper we propose an innovative method to integrate global and local feature space into a map function based on a novel distance function. A subset of train images denoted as exemplar-set are selected. The local and global distances are defined according to the images in the exemplar-set. Distances are defined such that they indicate the contribution of different semantic aspects and global information in each scene category. An empirical study has been performed on the 15-Scene dataset in order to demonstrate the impact of appropriately incorporating both local and global information for the purpose of scene recognition. The experiments show, our model achieved state-of-the-art accuracy of 87.47.