Applied Object Detection - HOW TO TRAIN YOUR D.R.A.G.O.N-SSD

Hey Everyone!

I'm happy to announce our next meetup! This time our speaker will be Dr. Ariel Biller from the research team of allegro.ai. Ariel will present allegro's recent work and improvements on top of SSD object detection in architecture as well as training strategies.

Big thanks to Taboola for hosting us in their place and sponsoring the event. Dan Friedman‏ will also share a short talk on the use of Computer vision in Taboola's Recommender system.

Agenda:
18:30-19:00 Gathering and networking.
19:00-20:00 Applied Object Detection/ Ariel Biller
20:00-20:05 Short break
20:05-20:30 Computer vision in Recommender system/Dan Friedman

Abstract:
DRAGON-SSD is our in-house implementation of SSD, developed at allegro.ai while seeking to provide a flexible, deployable prototype for detection and classification tasks for our customers. Based on the original caffe-SSD, DRAGON is focused on efficient transfer-learning and adaptations to unbalanced or small datasets.

In this session, I will give a general overview of the SSD meta-architecture and the new methods we have developed for DRAGON. Following two relevant cases of datasets for autonomous driving (KITTI) and smart-city mass surveillance (MOT), I will review the training-strategy built into DRAGON SSD and discuss how the same general considerations apply in both tasks achieving superior performance compared to SSD and YOLO.

About the speaker:
Ariel Biller received his PhD in (computational) chemistry from the Weizmann Institute of Science. Among his roles at the institute during the last decade, he was responsible for transforming scientific codebases to high-performance parallel computing software. Most notably, he participated in a joint academic-industry collaboration under the umbrella of the US Department of Energy towards exa-scale scientific computation, on which he gave a keynote at intel dev-con Tel Aviv 2017. Acknowledging the transformative power of deep-learning, he joined the research team at allegro.ai.

About allegro.ai:
allegro.ai is a pioneering deep learning computer vision platform that provides a complete product lifecycle management solution for AI development & production, starting with computer vision. The company’s platform simplifies the process of developing and managing deep
learning powered solutions - such as for autonomous vehicles, robotics, security cameras, logistics and others.
For more information, visit: http://www.allegro.ai

Taboola’s content discovery platform leverages computational models to match content to users who are likely to engage with it. Their content recommendation algorithm relies on a short description text and an accompanying image and integrates deep learning techniques from the field of NLP and computer vision. In this short talk, Dan will describe how Taboola have incorporated vision models into their recommendation system, outline the challenges faced in using transfer learning to predict click-through rate and review the technical aspects involved in incorporating these models into production.

About Dan:
Dan Friedman received his M.Sc. in computational biology from the Weizmann institute of science, specializing in machine learning models used to understand the genome. Since then he has been part of the Algorithmic group in Taboola, working on the core recommendations engine. Currently Dan leads a team that focuses on understanding image and text using deep-learning methodologies, and using these models to match recommendations to users based on their past preferences.