This class also comes on the tails of our Drafting and Drawing for Theater class the students have recently taken.
I am excited to share with you our latest advances in automatically recognizing street numbers visible in Street View images:
Note that this draft paper is under open review for the ICLR 2014 conference so feel free to endorse or criticize it and expect it to have rough edges in certain locations ;).
This work was done during 's internship at Google this past summer in close collaboration with and I. Past work on this topic has typically treated localization, segmentation, and recognition as separate tasks. We show in this paper that a deep neural network based probabilistic model, allows us to combine these three stages producing results superior to those reported in literature. Furthermore, we show that our results are comparable to human level performance on this task at high coverage on the public SVHN dataset. We also show that we improve on the state-of-the art for per character classification accuracy.
We have also trained our system on an internal dataset with tens of millions of samples coming from numerous countries. Our paper focuses on the transcription stage, but we are also able to detect street numbers from Street View at precision similar to human performance at really high recall as well. So far, we automatically detected and transcribed almost 100 million physical street numbers worldwide. Below is an image of automatically extracted street numbers in South Africa.
This achievement was possible due to long running collaborations between several teams within Google, specifically, with , and who were key contributors for the detection stage; as well as within several teams in StreetView for other parts of the project.
Finally I want to give a special thanks to my manager who provided me the best environment to work in.
So many times I have thought, "I remember thinking about this ages ago", or "didn't I make some notes on this ", the answer is more than likely, yet the notes I made are like the memory, long lost of forgotten. The real answe is ever again, and now I make...
- Konstrukshon LtdArchitectural Technologist, present
- Dow Chemical Co
One New Skill You Learned That Might Damn You More Often Than Help You
In one of our books, Julien Smith wrote about this concept where he said, "Should you be a hallway or a room?" He went on to explain the con
Meet the hexaflexagon. It's about to blow your mind.
Remember the first time you saw a Möbius strip (the ring-shaped surface with only one side) and it felt like your world had been turned upsi
Creative Time-Lapse of Kuala Lumpur Bounces Between Day and Night
If you have two minutes to spare, you've got to check out this time-lapse video by photographer Rob Whitworth. There are plenty of time-laps
Evgeny Morozov: The IGod: Steve Jobs’s Pursuit Of Perfection—and The Con...
Steve Jobs By Walter Isaacson (Simon & Schuster, 627 pp., $35). I. In 2010, Der Spiegel published a glowing profile of Steve Jobs, then
Nikon Unveils the Coolpix S800c, Its First Android-Powered Camera
After weeks of rumors and leaked photos, Nikon has finally officially announced its new Coolpix S800c compact camera -- the company's first
Autodesk - AutoCAD for Mac Services & Support - AutoCAD and AutoCAD ...
You want to know if AutoCAD for Mac and AutoCAD LT for Mac can be run on the new Mac OS X 10.8 (Mountain Lion) operating system.
HP Designjet & Dessau Institute of Architecture: Design for Venice B...
HP Designjet (find out more about HP Designjet products for architects here: http://budurl.com/7hcw) is teaming with SPAN Architecture and s
Konstrukshon CPD Weblog - Construction CPD for the Architectural Technol...
Architectural Technology,CPD,Architecture,Construction,building Technology
With Google's New Glasses, Your Head Is Your Smartphone
Google's Glass augmented reality system may be the smartphone (smartspecs?) of the near future, neatly linking our digital lives with our re