Past Meetup

Lou Kratz on Scaling Visual Search with Locally Optimized Product Quantization

This Meetup is past

36 people went

Details

Title:
Scaling Visual Search with Locally Optimized Product Quantization

Paper:
Locally Optimized Product Quantization for Approximate Nearest Neighbor Search.
Yannis Kalantidis and Yannis Avrithi.
2014 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2014.

Paper: http://image.ntua.gr/iva/files/lopq.pdf

Talk Abstract:

Approximate nearest neighbor (ANN) search is an essential technique for big data applications, and has only become more relevant as the scale of our data has increased dramatically in the past decade. At Curalate, we use ANN to power our visual search technology that enables our clients to identitfy apparel products in user-generated photos from social networks.

Kalantidis and Avrithi present an extremely fast and accurate ANN algorithm in their paper "Locally Optimized Product Quantization for Approximate Nearest Neighbor Search" (LOPQ). Their key contribution is a quantization method that minimizes distortion by leveraging Principle Component Analysis on small sub-vectors of the data. This technique enables fast and accurate searching over data sets of millions or even billions of items.

In this talk, I will present the LOPQ paper by Kalantidis and Avrithi, and follow it with a deep dive into how we have implemented it at Curalate to power our visual search technology. By leveraging LOPQ, we reduced our storage requirements by 90% allowing us to store an index of 31 million images in 1GB of RAM and search it in a only few hundred milliseconds.

About the Speaker:

Lou Kratz is a Research Engineer at Curalate, a Philadelphia-based startup that leverages social content to help their clients to sell more effectively online. He received his PhD in computer vision from Drexel University in 2012, and then got bit by the start-up bug in the best way. His passions include computer vision, machine learning, video analysis, and brewing beer when he can find the time.

This lecture will not assume any background in product quantization or approximate nearest neighbor search.

Directions:
When you arrive at 1900 Market St., present a valid photo ID to the security desk in the lobby. We will be located on the 8th floor. Once you exit the elevators on the 8th floor the entrance is very clear.

Food and drinks will be available at 6:30PM and the talk will start at 7PM.

Pizza and soft drinks will be provided by Curalate (https://www.curalate.com/).

All attendees are expected to follow the Papers We Love Code of Conduct (https://github.com/papers-we-love/papers-we-love/blob/master/CODE_OF_CONDUCT.md).

WeWork Blurb: WeWork is a community for creators. We transform buildings into beautiful, collaborative workspaces and provide the infrastructure, services, events and technology so our members can focus on doing what they love. WeWork currently has 111 locations in 29 cities across the world with over 70,000 members. Book a tour at wework.com now!