I succeeded to train my first cascade to detect car on the street from behind. Now I want to make it better since it still capture random trees and motorbike unless I adjust the parameter. I tried to add those specific trees and motorbike to the negative samples but the resulting cascade still detect them. Am I adding too little? I added 100 to about 1400. Also what other way could I optimize my cascade?
You mean 1400 negative sample? But if you ask me, it's not sufficient too. The negative sample is the one which your cascade should not detect. So please dont limit it with numbers, just add all the negative sample that hinders your perfect detection. So go for it. You mentioned something like your cascade detects trees even if you added it to negative samples. If it's the case, then add more number images containing that particular tree. And also add it's bark to the negative samples. Sometimes it might help you.