Improving eigenFaces by adding negative training images
This might seem like a dumb question but,
I i have 3 sets of images for 3 people in my database, and my recognition values are between 1 and 400, with threshold set at 400, if i add a couple of more people in my database, would that decrease the chance of getting false positives?
Now i currently still get someone who completely does not look like any of the 3 persons at 330, and still classifies as a known person. If i add a more diverse training set, would that be taken into account in calculating EigenFaces?
Thank you for reading.
imho, you need more images of the same person in the 1st place, 3 is not enough, try more like 10-20.
You misunderstood me, I have 3 sets of pictures comprising of 200 pics each. Sorry if I was unclear.
ooook, yes, indeed got that wrong. sorry forthe noise, then ;9
From what I remember eigenfaces always classifies as a known person. It just gives you the most similar. You need to use some sort of confidence value (like the one it returns or something else) to decide how likely the match is.
that's not the problem. The problem is that if i have a brown skin colored person in my training database, then a new untrained person shown to the webcam identifies nelson mandela as that brown person. president putin becomes any old person. etc.
this is really unacceptable and useless, so i'm wondering if i need to increase the variety of my training database to skew the eigenvalue threshold.
Yep variety in training data is the key to getting a better seperation. Also Eigenfaces might not be the best approach. Certainly compare the obtained accuracy to the Fisherfaces!
Hasn't anyone out there already done a comprehensive accuracy comparison of the different methods?
Well all comparisons are quite application specific in my experience. That is why you should always redo it for your own setup.
I'm using fisherfaces actually. Thanks for your input Steven. So you're saying add a person or persons (even if I don't want to recognise these) to improve my recognition rates of the others?
I had this idea to mitigate false positives by comparing known versus unknown matches but for that I would need to really track the faces, not just redetect them all the time per frame. Incidentally this would also save me a lot of cpu cycles.