evaluation based on continious variable

s_v1984 · Post by **s_v1984** » Wed Feb 04, 2009 7:13 pm

hi i want to classify skin pixel based on thier intesity in normalize rgb space i use only 2 channel for r and g as attached i was wondering if i train this by PC method how can i use it to make a decision(or get posteriori) if the pixel is skin given the r g ??
thanks

mark · Post by **mark** » Thu Feb 05, 2009 4:58 am

Yes, you can input this data to the PC algorithm, although it does assume that data is normally distributed. Why is skin always 1? After learning the model you can use inference to calculate the posterior probability distributions.

s_v1984 · Post by **s_v1984** » Thu Feb 05, 2009 11:15 am

Hi Mark thanks for quick response
the reason skin is 1 cos I inputted grand truth data(positive samples) where i am confident that r g gives skin colour i.e. I set them all to 1, when I try to train this it gives error where I cannot train discrete parameter with continuous one so i came up with two solutions:

1- discredited the parameters and use value to evaluate the skin posteriori value for given r and g, example attached.

2-remove skin column and then use continuous variable r and g by PC method to learn the network but in this case I am not sure how to evaluate the network as I cannot evaluate the values in genie. When I try to add arc between one of the continuous node it gives says it cannot connect arc, example attached

Appreciate any clue.

mark · Post by **mark** » Thu Feb 05, 2009 3:36 pm

You need to have negative samples as well, otherwise the algorithm thinks that all samples are positive and it won't work. After fixing this, learning should work.

s_v1984 · Post by **s_v1984** » Thu Feb 05, 2009 3:54 pm

well i have tried that already and it does not work i get the same error that i can not use the continuous and discrete parameter together, however i have managed to get rid of this error by changing one of the skin value to 0.1 instead of 0, then even if i get that how do i make inference to measure the skin outcome only if I have â€œrâ€ and â€œgâ€ value in my case.
Thanks a lot I really appreciate it

mark · Post by **mark** » Thu Feb 05, 2009 5:24 pm

I think the only option left for you is discretizing r and b and then learning. Have you tried that? (With including positive and negative samples.)

s_v1984 · Post by **s_v1984** » Tue Feb 17, 2009 1:47 pm

Hi thanks for you answer but my real intention by this example is to understand how I can train the continuous parameter and evaluate it by value node (so if I have r g then how can i get posterior without discreting the parameters).when I try to connect the arc between value node and continuous nodes it would not let me to do that I really appreciate if u can answer this

Thanks

mark · Post by **mark** » Tue Feb 17, 2009 5:40 pm

At the moment it is indeed not possible to connect continuous nodes to value nodes. We hope to add this feature in the future.

Snapel · Post by **Snapel** » Thu Feb 26, 2009 4:41 pm

Do you by any chance have an estimate on when this would be, I'd find this function very helpful, thanks.

mark · Post by **mark** » Fri Feb 27, 2009 1:04 am

Sorry, I don't really know. Hopefully somewhere in the next few months.