I was thinking about making a vector robot, similar to googles automonous
car... but then i thought - hey - if i just handle the simfirst i can plug him
straight into it!
i was thinking, maybe a stereoscopic htm, would be a far superior way to do
vision.
youd have to convert your depth map you compute it to some course mesh, (you
can remember the colours, if youd like colour playback)you do that (convert to
vector graphics) so its not voxels anymore, and now you can tween the recorded
states together, this will give you many more states as well as the states
provided by the exact record.
also, you pick some level of detail you are happy with, remember, its this
shape that it uses for similarity tests, so its kinda importantto get right.
thats basicly what it gives you. so imagine you rigged up a car with a stereo
camera, and you used this method.
go back home, and you playback, maybe it looks a little unrealistic when its
tweening missing information, but its better than if nothingwas there, maybe on
some error % you can refuse to tween then jump the player closer to the record.
and theres nothing stopping youfrom plugging it with fades of the colours
from the exact photos, for a little more closer to real, other than just
ordinary lines and rectangles.
but if you record so much, now youve got all the 3d tweens, that you wouldnt
have had just by recording colours... and it would be alot moreinteractive, and
"tweens of realistic physics" instead of exact realistic physics.
and thats something Jeff never mentioned. using geometry.
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
