Friday, January 10, 2014

Paris Machine Learning Meetup #7: Machine Vision: VMX, Atheer One and VLAD

From [1]

The Paris Machine Learning Meetup #7 will take place this coming wednesday Januray 15th. It will feature at least three presentations. Two of them will be made by people offsite and will focus on two crowdfunding campaigns that have a Machine Learning component. If everything goes well, we'll continue on experimenting with these off-site presentations. We've asked the meetupers to watch the videos first so they have questions ready for our speakers. Here is a draft program:

Abstract: The context of this presentation is visual search, with a specific focus on retrieving images similar to a query image. I will first discuss one corner stone of such large-scale systems: aggregation of local descriptors (typically SIFTs) into a fixed-sized image signature. I shall present "Vector of Locally Aggregated Descriptors" (VLAD) which offers a powerful alternative to popular "Bag-of-Word" (BoF) approach. Combined with an efficient indexing system, VLAD can lead to a memory footprint as compact as 16 bytes per image with good approximate search performance. I will then touch upon risks of visual information leakage in such image search systems, showing that human-understandable reconstruction of an image can be obtained from the sparse set of its local descriptors, and no other side information.

The Paris Machine Learning on LinkedIn:
To take part in the meetup, please register on, join the Paris ML group and go to this page.
The archives for the previous meetings can be found here.

[1] Reconstructing an image from its local descriptors by Philippe Weinzaepfel, Herve Jegou, Patrick Perez as featured here on Nuit Blanche

Organizers: Franck Bardol, Frederic Dembak, Igor Carron

No comments: