In this talk, I will first describe a family of metrics to compare images based
on their local descriptors. It encompasses bag-of-words and VLAD
representations, as well as matching techniques such as Hamming Embedding.
Making the bridge between these approaches leads us to identify successful
ingredients, and leads us to propose a match kernel that takes the best of
existing techniques by combining an aggregation procedure with a selective match
kernel. The representation underlying this kernel is then approximated,
providing a large scale image search both precise and scalable.

I will then present a recent work also related to large scale image retrieval,
where we consider the interest of oriented dense patches in the context of
aggregated representation like VLAD.