Building Recognition Using Gist Feature Based on Locality Sensitive Histograms of Oriented Gradients

Abstract

Locality sensitive histograms of oriented gradients based gist (LSHOG-gist) for building recognition is presented in this paper. Different from the traditional method which extracting orientation gist features by Gabor filters with only four angles, the proposed LSHOG-gist feature extraction method uses Locality sensitive histograms of oriented gradients of building images as orientation gist features. The LSHOG at each pixel is a multi-orientation histogram which is based on a whole building image. So, our LSHOG-gist is insensitive to noise such as non-uniform illumination or occlusion, and it has stronger texture description ability. Several experiments were conducted on the Sheffield Buildings Database, and satisfactory experimental results achieved, especially in the case of non-uniform illumination or occlusion.