Image indexing projects
Local descriptors
Together with Hervé Jégou, we worked on image indexing based on local descriptors with
Hamming Embedding, burstiness weighting,
and distance estimation, see our papers in
ECCV 2008,
CVPR 2009,
IJCV and this
research report. In this context, we gathered the
Holdiays dataset and benchmark. With
Jérôme Revaud also developed a specific method that is more appropriate for logo recognition, see our
ACM MM 2012 paper.
We developed the bigimbaz image indexing demonstrator, and a laptop version that just needs an external drive to store the image thumbnails (as raw data on a partition to avoid disk seeks!). Best demo award at RFIA 2010.
Global descriptors
We turned to global image representations like GIST (
CIVR 2009), bag-of-words representations (
ICCV 2009) and the VLAD representation, that was introduced in these
CVPR 2010 and
PAMI papers.
I developed INRIA's Fisher vector implementation.
Our GIST implementation (initially developed by Christophe Smekens) is here: lear_gist-1.2.tgz.
Nearest-neighbor search
Again with Hervé, we generalized local descriptor indexing methods to vector indexing in high dimensional spaces.
This lead to product quantization methods, and their extensions (IVF-ADC, etc.). They are described in our
PAMI and
ICASSP 2011 papers. In this context, a 1-billion vector dataset was introduced:
BIGANN.
Software
Most of the projects above use the Yael library, that is described in our
ACM MM 2014 paper, and can be
downloaded
on gforge.
INRIA sells licenses for the Obsidian (descriptor computation), Bigimbaz (image indexing), and PQ-codes (vector indexing) software libraries.
Misc
I also worked on color coding, see this
ACM MM 2011 paper.
I also combined unsupervised and attribute descriptors, see this CVPR 2011 paper.
I presented INRIA's image and video indexing technologies at the MPEG workshop on visual search in 2010. Slides.
