An analysis of SIFT object recognition with an emphasis on landmark detection

Abstract

In this thesis, I explore the realm of feature-based object. recognition applied to landmark detection. I have built a system using SIFT object recognition and Locality-Sensitive Hashing to quickly and accurately detect landmarks with accuracies ranging from 85-95%. I have also compared PCA-SIFT, a newly developed feature descriptor, to SIFT, and have found that SIFT outperforms it only particular data set. In addition, I have, performed a relatively extensive empirical comparison between Locality-Sensitive Hashing and Best-Bin First, two approximate nearest neighbor searches, finding that Locality-Sensitive Hashing in general performs the best.