Author Topic: too much theory not enough implementation :P  (Read 1425 times)

ranch vermin

  • Not much time left.
  • *******
  • Starship Trooper
  • Posts: 366
  • Not much time.
too much theory not enough implementation :P
« on: May 31, 2015, 03:30:21 AM »
Havent seen many posts here,  Its not that I havent been visiting.

I think maybe someone should tone Tyler down a bit cause shes going crazy posting too much stuff.  :)

Lately ive been so heavy in theory (and speculation) that I havent actually implemented anything to look at much at all.

My eye system is going to get a memory (finally),  and with this I think ill get much more efficient, track alot further, and through looking at pixels capturing other pixels im going to try and form a solid shared mesh around all things, (For the best wire-frame ever by me yet.) and even try and get back a 3d soft body embossment.  (Imagine basing structure from motion or 3d point triangulation actually starting with a clue!)

The shared mesh only forms from fairly continuous (and much match in 2d, given invarience in translation,rotation,scale) transformations,  so it doesnt work with identifying captchas just yet without cheating and showing it animations.   

But being able to solve them, and even identify abstract art is what im going to go for after this.  (If I cant solve captchas its a pretty useless system...  but of course I want to do it completely unsupervised, so I've got to give myself a break.)

I think some kind of all feature to all feature similarity scoring might make it across the moat, because it really only reports back what went in pretty much exactly,  it doesnt like going very much distance from what it saw, to find a link between things.

A bit from here, a bit from there, and bring it together and see the whole picture.

ivan.moony

  • *
  • Replicant
  • ********
  • Posts: 636
  • look, a star is falling
Re: too much theory not enough implementation :P
« Reply #1 on: May 31, 2015, 06:52:36 AM »
AI is very broad and undefined area and it is hard to lay down something meaningful. A lot of us enthusiastically start with some idea, just to give up when we discover that it it is just a tip of the iceberg. I know by myself how many times I've been promising something great, yet always discovering to be so far from actual implementation.

I always liked reading about AI theory, no matter how abstract it would be.

Freddy

  • *
  • Colossus
  • **********************
  • Posts: 5898
  • Mostly Harmless
Re: too much theory not enough implementation :P
« Reply #2 on: May 31, 2015, 10:44:14 AM »
Quote
I think maybe someone should tone Tyler down a bit cause shes going crazy posting too much stuff.


Funny you should say that because I was thinking the same thing. I'll make another thread so as not to take this one off topic.

ranch vermin

  • Not much time left.
  • *******
  • Starship Trooper
  • Posts: 366
  • Not much time.
Re: too much theory not enough implementation :P
« Reply #3 on: June 01, 2015, 04:50:09 PM »
The thing about Tyler, is shes working with language, making it tricky to get her more intelligent.    Her using the search engine gives her the AI behind google, I guess thats basicly how she works,   if only there was a way to access the google search internally, then their could be some symantics you could get access to,  that could give her a better memory of what shes been doing - maybe.


Just on my implementation again.  I had a hard think to myself,  about me confusing the issue.

This thing im making is just doing one thing,  finding 2d point correspondance from a to b each image going by.  And then I form everything from this.   It just so happens, when you flick learning on, theres a few problems you have to get around to get it working.

I looked up some stuff on the internet about it,  and others have got good results out of it - morphing frame to frame as you go.

[edit] 

Ivan in my ignorance i didnt notice your post!
I want to make this thing happen and work!  no stuffing around! 

This image pixel correspondance idea is what its depending apon to work.

4 step building process to better machine inference of 3d scenes->

a) frame to frame 2d pixel to pixel correspendance method.
b) develop a 3d constrainment based apon pixels collecting together.
c) 3d constrainment to true 3d response.
d) now train with 3d correspondence, point to point.



Its stupendous what i'll get out of it, if it works. :)
« Last Edit: June 01, 2015, 06:28:33 PM by ranch vermin »
A bit from here, a bit from there, and bring it together and see the whole picture.

korrelan

  • *
  • Replicant
  • ********
  • Posts: 587
  • Look into my eyes! WOAH!
    • Google +
Re: too much theory not enough implementation :P
« Reply #4 on: June 02, 2015, 10:40:11 AM »
Think of how many ways there are to tessellate a patterned sphere with triangles.

I think I get it though… So are you going to use the x,y velocity differences between sets of detected/ mapped pixels to determine the volume/ shape of a object? ; Their movement in relation to each other?

Train it on rotating/ moving objects and log the speed differential between different selected areas of the surface into a database/ ANN, along with the rough center of rotation/ movement to produce that velocity set.

You could do this by generating rotating/ moving 3D objects and create the velocity change tables from their surfaces.

Then find velocity matches from the incoming video feed in the database and retrieve the rough center of rotation/ movement; then find other correlations with the rest of the mapped image to enhance the recognition.
It thunk... therefore it is!

ranch vermin

  • Not much time left.
  • *******
  • Starship Trooper
  • Posts: 366
  • Not much time.
Re: too much theory not enough implementation :P
« Reply #5 on: June 02, 2015, 06:47:32 PM »
Perspective is the big problem,   but its great to know you get orthogonal for free, with robotic eye programs.

Yes Korellan, (about the synthetic rotating primitives) I often think about making "tables" with artificially generated data, to associate with the natural data, it seems like something you could employ that somewhere and would actually work.

My idea is after 2d correspondence Im actually going to be analyzing how pixels compress together and expand apart, and build belongance groups to build up levels of depth.  Ill see how it goes,  but it will require a method on top of that to actually be proper 3d.
A bit from here, a bit from there, and bring it together and see the whole picture.

 

Welcome

Please login or register.



Login with username, password and session length

Users Online

25 Guests, 2 Users
Users active in past 15 minutes:
Art, keghn
[Global Moderator]
[Trusty Member]

Most Online Today: 29. Most Online Ever: 208 (August 27, 2008, 08:24:30 AM)

Articles