Hello, good evening.

I am just starting learn about Hidden Markov Model to recognize motion picture. Assume in one frame I have m-dimension feature.

I want to ask, how to represent the sequence of data (which every frame has m-dimension feature) into the State Transition matrix and Emission matrix ? How many hidden state (S_1, S_2, S_h) that I need ?

I am trying to Google it but I didn't find it. It would be nice if someone can give the example too. Thanks :)

Update :

Ilustration :

alt text

Assume that one is for getting Probability for motion "A". And I have another model for motion "B".

asked Jul 09 '13 at 05:25

psuedobot's gravatar image

psuedobot
1111

edited Jul 11 '13 at 00:13


2 Answers:

I am not sure if I understand you. It would probably be better if you explain what the features are and what you are trying to do. I think your features are "tags" ie x , rather than the state "s"... if so then worth looking at this lecture http://www.cs.columbia.edu/~mcollins/loglinear.pdf

othewise if your m-dimensional feature is a state, then yes each (live) m dimensional feature combination is turned into a single tagged state [and you have a dictionary of states, state transitions ], so ecevything is the same

michael collins coursera lecture notes http://www.cs.columbia.edu/~mcollins/ explain and give peseudo code for hmms

answered Jul 10 '13 at 08:16

SeanV's gravatar image

SeanV
33629

hey, I tried to illustrate my model and update the question. Can you correct me if I do it wrong ?

(Jul 11 '13 at 00:14) psuedobot

HMMs are most frequently presented using a categorical distribution for p(x|s) (the emission probabilities) since this case is relatively easy to understand and what you would use when working with text. However there is no reason you can't use a more appropriate distribution here, like an m-dimensional Gaussian. Obviously you would need to also make the appropriate change to the way you calculate p(x|s) and update the emission probability parameters, but everything else remains mostly the same.

answered Jul 09 '13 at 08:38

alto's gravatar image

alto
60351124

so I must put my m-dimesional feature into 1 dimensional feature so I can present it into the state ?

also do you have any good resource to learn about HMM, I would like to try implement the HMM by hand so I can understand it better

(Jul 09 '13 at 14:44) psuedobot
Your answer
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.