ICA Training data

[Xue]

hi all   I want to use ICA to detect eye region (without eyebrow)    my questions are:   1- the number of training data will affect on the performance or not? for example I use 200 images only now if I increase to 1000 or 2000 images can the performance improve?   2- variation (lighting, view, direction, close, open,..) in the eye training images is important?   3- prepossessing of these training images will affect or not?   thank you so much for help

[Chris Tanner]

hi xue,

your questions are general machine learning questions, so the answers hold true for many applications.  i responded with in-line comments below:

On Thu, Dec 17, 2009 at 7:05 AM, Xue <humanf...@yahoo.com> wrote:

hi all   I want to use ICA to detect eye region (without eyebrow)    my questions are:   1- the number of training data will affect on the performance or not? for example I use 200 images only now if I increase to 1000 or 2000 images can the performance improve?

yes, the amount of training data will affect the testing performance.  typically, more training is better, but this relationship definitely isn't monotonic -- you can't expect to indefinitely get better performance by merely throwing more training data at it.  look up 'overfitting.' 

2- variation (lighting, view, direction, close, open,..) in the eye training images is important?

as with #1, having variation in your training data is important, well, especially if you will be testing future images that have variation.  think of your system as if you're trying to teach a kid to classify different objects.  whatever you will later test him on, well, you will help him do better if you teach him with data that is representative of that.  so, if you'll later ask him to distinguish b/w things that have much variety, then you should train him by showing him objects that have much variation.  

3- prepossessing of these training images will affect or not?

it could; it depends on what you do.  maybe your images are of really high resolution, and maybe not every single pixel (or some other feature) matters.  or, maybe every feature you have is useful for building your models.  preprocessing is a general term, so, there's never a clear answer, but typically SOME form of preprocessing is often needed. 

thank you so much for help

 

 

--

You received this message because you are subscribed to the Google Groups "Face Recognition Research Community" group.
To post to this group, send email to face...@googlegroups.com.
To unsubscribe from this group, send email to face-rec+u...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/face-rec?hl=en.

--
Chris TANNER
UCLA Graduate School '09 - C.S.
Florida Tech '06 - C.S., Math
www.chriswtanner.com photography
404-376-5817