Machine Vision on the Horizon: Vitamin D Launches Object Recognition for Video

readwriteweb.com

12 points·fjabre··9 comments

Static scenes, no apparent occlusion handling, poor object localization.

If I had to guess I'd say all they have going on in the demo is straightforward background-subtraction with blob tracking.

Maybe they're working on something more interesting, but haven't got a compelling demo yet. That being said, they're trying to tackle a very hard unsolved problem, and quite frankly their chances are low (using HTM is a bad sign.)

from what I've read, it seems that they are doing something to distribute a specially designed (tree structure) deep belief network across their clusters. Maybe the details more complicated (combining more properties of neuron?), and the results looks like that they just trained a "background-subtraction + blob-tracking" system.

Hmmm... that would be even worse. If you're right they not only have something that's trivial, but also opaque - as with almost all soft computing techniques.

What exactly is new here? I thought computers could identify and track an object in video already?

OK, let me see... the wide rectangles are cars and the tall ones are people. Is there anything else?

How is the new HTM method different from existing one, for example, Deep Belief Network?

Anyone download this yet? I'm dying to know what language it's written.

It looks like it's using Numenta's technology:

http://www.numenta.com/about-numenta/numenta-technology.php (link grabbed from the article)

Inspiring but kind of creepy... =)

Reply on news.ycombinator.com