Hi,
I had some cascade xmls with the trained samples but it has some problems on detecting and I am gonna repeat it with hard negative mining method. I am gonna collect the false positive results and add them into the my negatives.
I wonder if I should have used my old samples with the new false positives or change the old ones. Is there anyone having comment on this?
how large are your usual negative images ?
if they’re fairly large, hundreds of neg windows can be cropped from any of those,
so adding a few false positive windows to that pool will not make much of a dent, i’m afraid.
My negatives are as same as the positives. All pos and neg samples have 85x85 pixels. I think on creating the false positives on the result screen by capturing the result contours(the false ones only because the result contains the right object contours, as well.)
could you provide some positive images