Code for Vocabulary tree based image search
The "vocsearch" package contains source-code as well a pre-trained vocabulary tree and is therefore ready to use. The source-code is contained in a few classes only and does not need additional libraries. The vocabulary tree is trained on SIFT features created with Andrea Vedaldis SIFT implementation with is available at http://vision.ucla.edu/~vedaldi/code/sift/sift.html. The Matlab script "extractsift.m" shows how the store the extracted SIFT features in the format used in the "vocsearch" package. The "vocsearch" package includes an example program and example data. The example program shows image search on part of the UKY dataset (http://www.vis.uky.edu/~stewe/ukbench/).The example data consists of SIFT features extracted from the first 1000 images of the UKY dataset. The included vocabulary tree was trained on a large number of images from the internet.
The implementation is based on the publication:
"Scalable Recognition with a Vocabulary Tree", D. Nister and H. Stewenius.
(although it does not include the hierarchical scoring)
Download the package here: vocsearch.zip (76MB, source code + example data)
The source code was used for our publication:
Friedrich Fraundorfer and Changchang Wu and Jan-Michael Frahm and Marc Pollefeys. Visual Word based Location Recognition in 3D models using Distance Augmented Weighting. In Fourth International Symposium on 3D Data Processing, Visualization and Transmission., 2008. (PDF)
Watch the VIDEO to this paper!