Fast Landmark Indexing Using Connectivity between Visual Words

  • 林 芳君

Student thesis: Master's Thesis


Bag-of-visual-words (BoVW) model has been widely adopted in the multimedia and the computer vision community e g image classification object recognition and landmark retrieval In BoVW framework visual words are built by clustering SIFT descriptors extracted from images Each image is represented as a histogram containing the count of each visual word Due to its ignorance of the spatial context information the performance of object retrieval will decrease dramatically To solve this the objective of this thesis is to propose a novel geometric verification that captures the relative spatial information of visual words and then constructs the structure representation of a landmark to improve the standard BoVW model Moreover to get better results the modified TF-IDF technique is used to calculate the significance of visual words for obtaining a more accurate weighted feature vector of image In the experiment two image datasets are used for the evaluation i e Oxford Building 5K and Paris 6K dataset The experimental results show that the proposed approach has better performance compared to other methods with spatial models with efficient post-processing modules
Date of Award2016 Feb 1
Original languageEnglish
SupervisorShen-Chuan Tai (Supervisor)

Cite this