[ranch vermin]

(look at it fullscreen to see the segments.)



Inspired by the predator computer vision project, this is my next video net for 2015, putting a spin on it.

I had the idea a while back, predator has been around since 2010, its just when I first budgeted making this thing, it was so computational I gave up on it with my teeth chattering at the insane astronomical numbers of things you need.

However I dont know if it will work, Im just uploading this because it has an interesting retina idea, Ive never done before.
It resembles something youd think maybe an insect with a compound eye would have.

Every pixel has a slightly offsetted image inside it. (I offset 1 every tile) (this is only a 128x128 resolution picture... but its jumped up to actually run this thing - a picture in every pixel.) I then run a mini net per pixelchunk/tile, in parallel... as aposed to one big net, Maybe akin to using gpu over cpu.
This is so, if one pixel learns something, then all pixels know it.
No matter where something appears (in 2d), it is the same position to the network.
I cant use the edges as well as the centre, because theyre gathering off the edge of the screen, more the more closer they are to it.
And if I want spin invarience, then I cant use the corners of the minitiles, only the contained circle can be used, and I can only use an odd number radius, to have a real centre pixel. makes for a cylindrical type network topology.

Its not for the processing power tight fisted, and Im a little worried that itll just halt the computer with it all on... its effectively running a whole net per pixel, its not as easy task, and requires downgrading the capacity and resolution on a hefty gpu to work at all.

Its pure rgb based tracking, but you could hook it up with a kinect and get a nice duo.  They would definitely help each other,  kinect supplying the depths,  the predator doing the surface tracking.

[Snowman]

I don't know exactly what your doing with this, but I'm interested in finding out. Are you looking into computer vision or is this meant to be an artistic approach? You've obviously given this a lot of thought.

[ranch vermin]

THIS IS GOING BACK AGES NOW!
its a key maker instanced per pixel.   which i then bruteforce the near area and find the location of the pixel in the next frame.