Hello everyone
I'm currently working on classifying different 3d signs in a sign language project and i'm using kinect for extracting 3d joints locations for both hands and this is considered the sequence i'm using to train the HMM.
Although the signs i'm trying to classify are a little bit different from each other, HMM isn't performing as expected and it misses some signs and mispredicts the sign, currently i'm trying to optimize its performance and don't know where to start
so my questions are (i need any help so if you can only answer one of these that would be great :) ) :
1- How many samples per sign should i train the hmm?
2- should the samples include variations of the same sign with multiple signs representing the average movement of the sign ? should i train the hmm with different signers ?
3- what are your suggestions for the preprocessing steps that i should do before learning and classifying for the signs itself, currently i'm subtracting the joints xyz from a stable joint the shoulders center ?
4- do you suggest subtracting the mean of the signs ? dividing by the standard deviation? i know that this will make some signs look identical while they're not actually the same.
5- do you expect that if i added the xyz of the elbows to the feature vector that would be better for training the hmm as some signs need elbows movement and others don't?
6- Should each model have different parameters ? i.e hidden states numbers or the whole classifier should have the same count for states
I'm very thankful in advance :)
Salam, Omar,
I do not answers for ll questions. I am not an expert but I am just trying to help. If you can discuss with someone who is specialized in Robotics, he will be of high value for such decisions.
1- How many samples per sign should i train the hmm?
No exact number, use the just enough number *125%. Allocate 25% for testing. Take care to represent all variations you need, (man, woman), different ages,..etc.
2- should the samples include variations of the same sign with multiple signs representing the average movement of the sign ? should i train the hmm with different signers ?
From my point of view, the answers are: Yes, Yes
3- what are your suggestions for the preprocessing steps that i should do before learning and classifying for the signs itself, currently i'm subtracting the joints xyz from a stable joint the shoulders center ?
I suggest using the Kinnect model, referring for each sign relative to its axis of rotation, translation. not to the global fixed point. In the same time, using coding for the preconditions of the movement.
4- do you suggest subtracting the mean of the signs ? dividing by the standard deviation? i know that this will make some signs look identical while they're not actually the same.
I think no, some signs are known by their magnitude.
5- do you expect that if i added the xyz of the elbows to the feature vector that would be better for training the hmm as some signs need elbows movement and others don't?
6- Should each model have different parameters ? i.e hidden states numbers or the whole classifier should have the same count for states
How to train Hidden markov model for 3d gesture recognition?. Available from: https://www.researchgate.net/post/how_to_train_Hidden_markov_model_for_3d_gesture_recognition [accessed May 30, 2015].
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation