suneelb's profile - activity

Hi!

I am trying to train a cascade classifier to detect a logo, and I keep getting stuck on stages. I started off with 200 pos and 700 negative samples, and this wouldn't even train past stage 2 (which also stopped on the same POS count line below, but for stage 2). When I found more samples (2600 pos and 5000 neg), I made it to Stage 4 before the infinite loop started again. It is usually stuck on this line:

===== TRAINING 4-stage =====
<BEGIN
POS count : consumed   2692 : 2696

Here is the command I am running:

opencv_traincascade -data classifier -vec total_vec.vec -bg bg_real_deal.txt -numPos 2692 -numNeg 5138 -numStages 10 -precalcValBufSize 4096 -precalcIdxBufSize 4096 -featureType LBP -minHitRate 0.998 -maxFalseAlarmRate 0.5 -mode ALL -w 100 -h 80

Does it matter that I am using LBP? Should I change the number of positive and negative samples? What is the ideal ratio of pos to neg? I know CPU isn't a concern, as I am running this training on an AMI. Any other tips or fixes?

Thanks!

Could you expand on step #6 and #7 and how you did that in the code (meaning how you combined the svmlight model vectors into a single vector)?