| View previous topic :: View next topic |
| Author |
Message |
mik
Joined: 10 Jan 2007
Posts: 26
Digg It |
 Posted: Wed Apr 25, 2007 6:58 am Post subject: |
|
|
How is your net build up? You have the values for the 3 axes as one input, but how to you resolve the sequential character?
I'm not so deep in Machine Learning, but i would guess that HMMs are a quite good approach for sequence classification, as they are also commonly used for speech recognition. Btw, a good introducing paper for ML ond Sequential Data is:
Machine Learning for Sequential Data: A Review
I think i will wait until something happens at the GART project, and that put some effort into testing this out. Based on the last update dates in the wiki this project is still active. |
|
| Back to top |
|
 |
mik
Joined: 10 Jan 2007
Posts: 26
Digg It |
| Posted: Wed Apr 25, 2007 7:09 am Post subject: |
|
|
| wic wrote: | | I read somewhere about using "acceleration deltas", whatever that means. |
Acceleration deltas....hmm... maybe instead of using the absolute acc. values just calculate the delta between the last and the current value, like:
delta x=x(t)-x(t-1) |
|
| Back to top |
|
|
wic
Joined: 26 Mar 2007
Posts: 57
Location: Sweden
Digg It |
| Posted: Wed Apr 25, 2007 9:38 am Post subject: |
|
|
| Quote: | | How is your net build up? |
I'm using 600 neurons as input, some number of hidden neurons and four neurons in the output layer. I've limited a motion to 2 seconds which gives a maximum of 200 points (200x3 = 600 inputs). I've trained the ann with 4 different gestures (circle, square, two, three) hence 4 outputs.
| Quote: | | but how to you resolve the sequential character? |
Dunno. I'm not even sure this is a sequential problem as described in the paper, or at least I dont understand what is so sequential about it. I'm treating this more or less as OCR but in 3D  . So please elaborate.
| Quote: | | delta x=x(t)-x(t-1) |
Yup, I tried that. Made no difference. I dont get why using deltas should give any better result. The only thing I can think is that you are sort of calculating the derivative of acceleration (ie the 'jerk') and that may say something of the movement. |
|
| Back to top |
|
|
wic
Joined: 26 Mar 2007
Posts: 57
Location: Sweden
Digg It |
| Posted: Wed Apr 25, 2007 11:02 pm Post subject: |
|
|
A short update on this.
I've managed to get the neural net working and it's quite good. Not aiLive-good, but it can recognize between four guestures right now.
I really think some kind of pre-processing of the acceleration data would help things, but I dont know what what could be.
At least this is a proof of concept that it can be done. |
|
| Back to top |
|
|
TiagoTiago
Joined: 20 Jan 2007
Posts: 710
Location: Brasil
Digg It |
| Posted: Thu Apr 26, 2007 12:50 am Post subject: |
|
|
| wic wrote: | ...
I really think some kind of pre-processing of the acceleration data would help things, but I dont know what what could be.
... |
just plain smoothing out isn't doing it?
_________________
please put the scripts on the wiki so they dont get lost as new stuff is posted!
phpBB doesnt like me,somtimes it will forget to warn me about new replies to threads I asked it to,if you see a thread I should have responded, could please email me? |
|
| Back to top |
|
|
mik
Joined: 10 Jan 2007
Posts: 26
Digg It |
| Posted: Thu Apr 26, 2007 7:11 am Post subject: |
|
|
Nice! So how to you decide which data you take? To you always take the last 200 inputs as a input sequence? So for each t you process all t, t-1,...,t-199 values? Or to you start your gestures by pressing a button?
One challenge i see is, that each of your gestures has to be exactly 2 seconds long to be detected best. With that fixed kind of window, it might be an approach to always use the full 200 values by interpolating the values up when the gesture is to short, or down, if the gesture is to long. This for sure depends, on how start and end of the gestures is defined/detected.
I programmed some java stuff, that tries to find out if currently a gesture is performed or if it is just minimal jitter movements that come out of the wiimote. But than I decided to wait for the GART.
Anyway, Congrats that you made it. |
|
| Back to top |
|
|
mik
Joined: 10 Jan 2007
Posts: 26
Digg It |
| Posted: Thu Apr 26, 2007 7:15 am Post subject: |
|
|
| Btw, in which language are you coding? And are you using a specific machine learning toolkit. I looked at Joone and Jahmm for Java. |
|
| Back to top |
|
|
|
FAQ
Search
Memberlist
Usergroups
Register
Profile
Log in to check your private messages
Log in
