|
|
Post by Pttg on Apr 14, 2017 19:48:01 GMT
Using a generous definition of "anything" for this discussion, let me know if this thread needs to be moved. I installed Pix2Pix on my computer, which isn't terribly impressive but nonetheless now lets me use my crappy 740m Nvidia as a vision processing neural network. All it can do is turn one image into a different image. Now I can't think of what to do with it. Or rather, I can think of too many things. I could take pictures of faces from a database and teach the computer to draw left halves when given right halves. I could take pictures of road maps and train the computer to turn that into satellite images. I could teach the computer to turn all fonts into arial when given a bitmap image.... So I can't decide what to do. Any ideas? The limitations include the need to have between 300 and 2000 source images, and the inputs and outputs I _think_ need to be 256x256. Ideally, there should be some connection that a normal filter couldn't discern. Also, ideally it won't be beyond a trivial level of python fiddling to implement. |
|
|
|
Post by bigbombr on Apr 14, 2017 20:18:30 GMT
Perhaps learn it how to make textures? Example: you give the texture of a wooden slope, and it gives you the texture of a wooden cube. This would massively speed up game development.
|
|
|
|
Post by Durandal on Apr 14, 2017 20:54:03 GMT
Teach it to be a benevolent overlord please. And give it a good mead brewing recipe for our ration alotments.
|
|
|
|
Post by Enderminion on Apr 14, 2017 21:54:53 GMT
teach it to make beer
|
|
|
|
Post by Pttg on Apr 14, 2017 22:31:17 GMT
All good ideas, but the network I have can only do picture-to-picture transformations.
|
|
|
|
Post by someusername6 on Apr 15, 2017 16:20:26 GMT
Create edges2spaceship, like edges2cats but using spaceships instead of cats. |
|
|
|
Post by argonbalt on Apr 15, 2017 16:38:36 GMT
You know it suddenly occurred to me that such a system might be interestingly useful in target acquisition, if you had a reference of numerous ship silhouettes you could quickly run them through an observational scan and see if they can pick up any targets?
|
|
|
|
Post by goduranus on Apr 20, 2017 3:26:44 GMT
I myself is dabbling with tensorflow, to analyze "something", don't know what yet, lol.
|
|
|
|
Post by Pttg on Apr 20, 2017 5:35:03 GMT
I've actually got an idea: have a database of chessboard photos. Train it with before and after pictures where white makes a wise move.
After long enough, it's almost a chess AI. Albeit one that occasionally recommends blurring all the pieces into an unrecognizable mass.
|
|
|
|
Post by argonbalt on Apr 21, 2017 0:49:57 GMT
I've actually got an idea: have a database of chessboard photos. Train it with before and after pictures where white makes a wise move. After long enough, it's almost a chess AI. Albeit one that occasionally recommends blurring all the pieces into an unrecognizable mass. You know this instantly reminded me of one of the best scenes from The Black Hole Where they find the main ship the USS Cygnus and have to scan through matching silhouettes until they get a lock.  Essentially if you could snapshot a ship, silhouette it and scan it against its background radiation you could visually target lock it based on a 2 or 3d gridded silhouette signature. The hard part would be the fact that most of the precise targeting requires a smaller chunk of that (usually a grey turret orb or a black rectangular radiator). Granted doing that might be as simple as asking the guns to *ahem* trim down the silhouette so to speak. |
|
|
|
Post by nerd1000 on Apr 21, 2017 8:57:23 GMT
I've actually got an idea: have a database of chessboard photos. Train it with before and after pictures where white makes a wise move. After long enough, it's almost a chess AI. Albeit one that occasionally recommends blurring all the pieces into an unrecognizable mass. You know this instantly reminded me of one of the best scenes from The Black Hole Where they find the main ship the USS Cygnus and have to scan through matching silhouettes until they get a lock. Essentially if you could snapshot a ship, silhouette it and scan it against its background radiation you could visually target lock it based on a 2 or 3d gridded silhouette signature. The hard part would be the fact that most of the precise targeting requires a smaller chunk of that (usually a grey turret orb or a black rectangular radiator). Granted doing that might be as simple as asking the guns to *ahem* trim down the silhouette so to speak. Maximilian! |
|
|
|
Post by argonbalt on Apr 21, 2017 14:08:56 GMT
You know this instantly reminded me of one of the best scenes from The Black Hole Where they find the main ship the USS Cygnus and have to scan through matching silhouettes until they get a lock. Essentially if you could snapshot a ship, silhouette it and scan it against its background radiation you could visually target lock it based on a 2 or 3d gridded silhouette signature. The hard part would be the fact that most of the precise targeting requires a smaller chunk of that (usually a grey turret orb or a black rectangular radiator). Granted doing that might be as simple as asking the guns to *ahem* trim down the silhouette so to speak. Maximilian!
you'll suffer till the end of time, enduring tortures, most of which rhyme,
trapped forever here in Robot Hell!''
Seriously that movie is al over the place so much it is not even funny. |
|
|
|
Post by nerd1000 on Apr 22, 2017 5:37:22 GMT
''you'll pay for every crime, knee-deep in electric slime,
you'll suffer till the end of time, enduring tortures, most of which rhyme,
trapped forever here in Robot Hell!''
Seriously that movie is al over the place so much it is not even funny. It's unusually dark and loaded with religious themes for a Disney movie. I think the original intention was for it to be a scary psychological type of film, but the success of Star Wars influenced them to add more 'friendly robot' type characters and some more light hearted scenes that didn't really fit in (as an example, the robot shooting gallery competition scene). |
|
|
|
Post by Enderminion on Apr 22, 2017 11:10:19 GMT
yeah, "robots"
|
|
