Prev: COM Automation
Next: NLMIP
From: Ali Harja on 10 Aug 2010 23:57
dear sirs,

I have currently started a project with school, concerning prediction of the path of 2d movement of fish. I have a set of pics with 1 fish in a very large tank i.e. each frame has set of elements(co-ods of cneter x , y, area, anglel) of the fish: which features may also be useful for track?

so i want to predict position of fish in 8th pic given data of previous 7.

I have data = 4(properties) x 100(pics).

I am unclear about how shall I train the I-H-O network use nftool.

I nned to prepare input and target vectors.

I tried the following:

input=[pic1 pic2 pic3... pic7]
target=[pic8]
where input picn is data(:,n),
and target picn is data(1:2,n); i.e just coords

but did not like dimensions, need a output per input?
i belivve the network should have 4 input nodes(each feature) but in this case it gives me 7(amount of pics). nftool seem to need number of picture samples the same for input and target. So I am unsure how to train net to get one set of values from the past 7.

so instead, shall I use?
input=[pic1 pic2 pic3....
target=[pic8 pic9 pic10....
although i dont think this use 7 past pics, only the n-7...

shall i use a sort of delay? or a structure of input and target i have not thought?

looking forward for any suggestion
Ali H
From: Greg Heath on 12 Aug 2010 04:25
On Aug 10, 11:57 pm, "Ali Harja" <jjsw...(a)hotmail.co.uk> wrote:
> dear sirs,
>
> I have currently started a project with school, concerning prediction of the path of 2d movement of fish. I have a set of pics with 1 fish in a very large tank i.e. each frame has set of elements(co-ods of cneter x , y, area, anglel) of the fish: which features may also be useful for track?
>
> so i want to predict position of fish in 8th pic given data of previous 7..
>
> I have data = 4(properties) x 100(pics).
>
> I am unclear about how shall I train the I-H-O network use nftool.
>
> I nned to prepare input and target vectors.
>
> I tried the following:
>
>  input=[pic1 pic2 pic3... pic7]
>  target=[pic8]
> where input picn is data(:,n),
> and target picn is data(1:2,n); i.e just coords
>
> but did not like dimensions, need a output per input?
> i belivve the network should have 4 input nodes(each feature) but in this case it gives me 7(amount of pics). nftool seem to need number of picture samples the same for input and target. So I am unsure how to train net to get one set of values from the past 7.
>
> so instead, shall I use?
> input=[pic1 pic2 pic3....
> target=[pic8 pic9 pic10....
> although i dont think this use 7 past pics, only the n-7...
>
> shall i use a sort of delay? or a structure of input and target i have not thought?
>
> looking forward for any suggestion
> Ali H

sizeInput = [28 93]
sizetarget = [2 93]

Hope this helps.

Greg
 | 
Pages: 1
Prev: COM Automation
Next: NLMIP