v0.9, Andrea Vedaldi, University of Oxford.
VisualIndex is a MATLAB and VLFeat-based program to index and query visually a collection of images. This system obtains near state-of-the-art performance on the Oxford 5k benchmark. Specifically:
demo_oxford5k- TF-IDF + spatial verification index with 83.9% mAP.demo_oxford5k_baseline- Direct image matching (no index) with 87.7% mAP.
This implementation is very compact and simple, useful for education and research.
The home of this package is on GitHub (https://github.com/vedaldi/visualindex). The code requires MATLAB 2009B onwards and requires installing the VLFeat library in the vlfeat/ subfolder (VLFeat is configured as a submodule of the git repository, and can also be downloaded at http://www.vlfeat.org).
Using the code is simple. Given a cell array of paths to images
images, start by initializing the index structure:
index = visualindex_build(images) ;
This will estimate a visual word vocabulary, but will not add the images to the index yet. Use options such as 'numWords', to specify how many visual words to use in the creation of the index (see the inline help for details). To add images to the index use:
visualindex_add(index, images, ids) ;
where ids is a vector of numeric image IDs (used as a convenience
instead of image names to refer to images later). To query the index
use
[ids, scores] = visualindex_query(index, image)
where image is either an ID of one of the images added to the index,
a new image, or a path to a new image. Use options such as
geometricVerification to control aspects of the query process. Use
the box option to restrict the search to a box. The function returns
a list of image ids and matching scores by decreasing confidence.
To visualize the top matches use:
[ids, scores, result] = visualindex_query(index, image, path) ;
visualindex_show(index, result) ;
To run the demo, make sure that VLFeat is contained in a subfolder of the VisualIndex package directory and that it is compiled. Then, run MATLAB and issue
> demo_oxford5k
to automatically download the Oxford 5k building dataset and start
running demo queries. The indexing process takes a while, but can
accelerated significantly by distributing the work on multiple cores
or computers by opening a matlabpool (see matlabpool in MATLAB).
Furthermore, VLFeat can take advantage of multiple core CPUs if MATLAB
is not started with the -singleCompThread option.
- 0.9 - Significant performance improvements
- 0.1 - Initial release