Understand instead of Memorize
Depth is crucial in understanding
I think computer vision (based on machine learning) will improve a lot (more) when we’ll find ways to understand a scene or even better a sequence of (related) images, like in a real time video stream.
As fascinating the results presented in the article are, they are the result of efficiently memorized facts, computed through training on a lot of images, that still represent a very (very) small fraction of all the real life situations.
I played recently with SSD Tensorflow implementation, and I’m really impressed by the accuracy (and speed) of the detection, as you can see here in this video (similar accuracy on videos recorded from my car), easy to use source code in video description, if you’ll like to reproduce the results:
SSD Tensorflow based car detection and tracking demo for OSSDC.org VisionBasedACC PS3/PS4 simulator
I’m very interested in techniques that understand the scene, so we can be 100% (like humans with good vision) sure, especially in self driving car space, anything less could be devastating for some.
I feel that without depth information, from stereo or multi camera setups, it will be harder to achieve full autonomy on all streets in the world in next 5 years.
That’s why I started to collect ideas on how to build an affordable smart camera, that will allow everyone to collect, analyze, test and help us achive that safely.
Thanks for the nice article!
If anyone is interested to apply these techniques in Self Driving Cars and Robotics, join our efforts in Open Source Self Driving Car Initiative:
You can read more about the OSSDC Initiative here:
