Quest for Real time Object detection in Mobile Browsers

  • Tuesday, Aug 7, 2018

If you are following Kepler work closely for past few months a certain emerging trend is going on, we are extensively seeking out opportunities for developing solutions based on Augmented Reality. Just take a look at
Some of our past works dinner tonight ar app for recipe recommendation and Reimagining The Tour Guide
Both of these work has one thing in common. Here we are using Native Augmented Reality Platforms specifically iOS ARKit.
One of the challenges of developing Native Augmented Reality / Machine Learning based app is customer Reach.
It is very difficult to convince people to install a new App. Here Mobile Browser-based applications shine.
The first Step of Development of Interactive ML-AR application is Real-time Object detection on Mobile.
So how can we develop them? So we set out on the mission to find the secret solution.




So how do we start you may ask ?

We started with using Tiny YOLO object detection model in darknet network, pretty Dark path have we choson :)

Converted it into Tensorflow js model and then started detection.

This is how detection looks like.



A problem, It takes 800 ms for each frame we need a solution.