ºÝºÝߣ

ºÝºÝߣShare a Scribd company logo

Computer Vision Meetup March: How we built snapscreen

•Download as PPTX, PDF•
•
Anyline
Anyline

Talk from Thomas Willomitzer at the Computer Vision Meetup March Edition about "How we built snapscreen"

1 of 10
Download to read offline
How we built Snapscreen From idea to scalable system
Initial idea • Watched Superbowl game in Silicon Valley in 2013 and saw Shazam enabled commercial on TV • Worked at Jumio as CTO at the time having lots of exposure to computer vision • Added 1+1 to come up with the idea for Snapscreen • Initial idea was to take a photo with any phone and send photo to email address. In reply email send contextual information about program you are watching
First steps • Used OpenCV to implement MVP to proof the idea can work • Did some homework on precision and scale metrics • Did some homework on cost both in production of technology and operation of technology for thousands of TV stations and millions of users • Seemed doable – so we went ahead
The first year • Decomposed the problem into smaller problems: screen detection, image matching, TV data gathering and storing • Building of data sets for development and test of algorithms for each smaller problem • Solving of each of the smaller problems
The technology behind Snapscreen • CI – nearly CD – built on Jenkins, Artifactory, Chef and Docker • Highly scalable backend built on Java using state of the art Spring Cloud • Scaling built on Digital Ocean and Zabbix • Persistence layers built on top of Postgres with CitusDB and MongoDB • Computer Vision layers built with C++, Neon, AVX2. All driven by CMake and CTest
Computer Vision Meetup March: How we built snapscreen
Computer Vision Meetup March: How we built snapscreen
The product offering • The vision is to automate the search for anything on TV • For TV Apps this is identifying the program and deliver SERPs (search engine result pages) for program title, actor names, episode information etc … • For Betting Apps this is finding the right game and bet immediately – keep the impuls • For Sports Apps – It’s a hybrid between TV Apps and Betting Apps. We find the right game in the app – and also deliver SERPs for Teams and Players
Lets see Snapscreen in action • Show video https://www.snapscreen.com/intro-video • Show live demo (and hope it works) • Assuming the demo worked – who wants to try?
Computer Vision Meetup March: How we built snapscreen

More Related Content

Computer Vision Meetup March: How we built snapscreen

  • 1. How we built Snapscreen From idea to scalable system
  • 2. Initial idea • Watched Superbowl game in Silicon Valley in 2013 and saw Shazam enabled commercial on TV • Worked at Jumio as CTO at the time having lots of exposure to computer vision • Added 1+1 to come up with the idea for Snapscreen • Initial idea was to take a photo with any phone and send photo to email address. In reply email send contextual information about program you are watching
  • 3. First steps • Used OpenCV to implement MVP to proof the idea can work • Did some homework on precision and scale metrics • Did some homework on cost both in production of technology and operation of technology for thousands of TV stations and millions of users • Seemed doable – so we went ahead
  • 4. The first year • Decomposed the problem into smaller problems: screen detection, image matching, TV data gathering and storing • Building of data sets for development and test of algorithms for each smaller problem • Solving of each of the smaller problems
  • 5. The technology behind Snapscreen • CI – nearly CD – built on Jenkins, Artifactory, Chef and Docker • Highly scalable backend built on Java using state of the art Spring Cloud • Scaling built on Digital Ocean and Zabbix • Persistence layers built on top of Postgres with CitusDB and MongoDB • Computer Vision layers built with C++, Neon, AVX2. All driven by CMake and CTest
  • 8. The product offering • The vision is to automate the search for anything on TV • For TV Apps this is identifying the program and deliver SERPs (search engine result pages) for program title, actor names, episode information etc … • For Betting Apps this is finding the right game and bet immediately – keep the impuls • For Sports Apps – It’s a hybrid between TV Apps and Betting Apps. We find the right game in the app – and also deliver SERPs for Teams and Players
  • 9. Lets see Snapscreen in action • Show video https://www.snapscreen.com/intro-video • Show live demo (and hope it works) • Assuming the demo worked – who wants to try?