stefan wrote:Hello everyone,
I'm also working with P300 picture-based speller. I've changed the letters into icons but the problem is that the icons do not increase on size or change the colours when they flash, not like the letters. Thus I got low accuracy of 57.9% using the icons
comparing with 83.33% using the letters. Is there any way to change the colour or incease the size of the flashing icons to follow the same effects of the letters?
I'm not sure how to do this, but maybe the developers can help you with that. Anyway did you try the p300-magic-card scenario that comes with OpenViBE? This scenario has some kind of effect that hides the picture, which maybe could enhance performance?
The accuracy that you are mentioning is the accuracy given when you trained the classifier I guess, not the one of the online testing phase?
stefan wrote:Also, I got this error during the online testing: "[WARNING] <Algorithm::LDA classifier - GPL> Feature vector size 57 (0x39) and hyperplane parameter size 229 (0xe5) does not match"
How can I avoid this problem?
cheers
This could be caused by several things. Either you did not train the classifier and used a classifier that was trained on feature vectors of 229 dimensions, while you only extract 57 features during online testing. Or you forgot to change some settings in the online scenario to match your settings of the training scenario (or vice versa). For example if you concatenate three channels of 60 samples, then you have a feature vector of 180 dimensions and your trained hyperplane of the LDA classifier should also count 180 dimensions (plus some additional bias term which we forget about for now). If of course you train your scenario with three concatenated channels of 60 samples and you only select one channel (or one spatial filter for the XDAWN algorithm) then you only have a feature vector of 60 dimensions during the online testing which is is a mismatch with the dimensions of the hyperplane.
Hope this helps,
Best regards,
Dieter Devlaminck