Aggregating local descriptors into a compact representation
We provide some Matlab package for the method of our
CVPR'2010 paper,
which describe a large scale image indexing system with a very compact image
representation (an image is typically represented by 16 to 40 bytes).
This package is Matlab implementation of our approximate nearest neighbor search
with a compact memory representation of the index, which is described in
this paper.
The Yael library provides
efficient implementations of computationally demanding functions, such as kmeans and exact k-nearest neighbors search (used, e.g.,
to assign descriptor to visual words based on a k-means codebook).
Download our own C implementation of GIST (by Matthijs Douze and Christophe Smekens) that we developped for our CIVR'2009 paper: "Evaluation of GIST descriptors for web-scale image search".
Anti-sparse coding
Anti-sparse coding is a technique that is opposite in spirit to sparse coding:
Instead of trying to concentrate the signal on a few components, the objective of these so-called "spread representations" is
to use almost all the components so that all have a comparable contribution to the final reconstruction.
This
link
provides the Matlab package that reproduces some results of our paper
"Anti-sparse coding for approximate nearest neighbor search"
(Jegou, Furon and Fuchs). It includes the L-infinity solver used to produce
the spread representations.
Babaz: audio indexing system for video copy detection
Audio search system for copy detection available
here.
