1

I'm a perfect newbie with cascade training. I managed to use the Cascade Training Gui software ( http://amin-ahmadi.com/cascade-trainer-gui/ ), it seems to work okay, except that the XML classifier produced never worked (I currently use processing, but I guess it's not the problem). The process is terribly slow for each tryout. I'm beginning to think I miss some information...

So my question is: what positive images should I use precisely? I generated hundred of images like these:

the positive images I used

Does it seem like the good way to do it?

rollstuhlfahrer
  • 3,988
  • 9
  • 25
  • 38
  • 2
    afaik haar and lbp best work for textured objects. Shape should go better with gradient based classifiers like HoG but not sute, maybe haar/lbp could handle those, too... – Micka Apr 29 '18 at 22:13
  • Thank you for this answer. Can you give me an advice regarding a tool I could use to work with HoG ? Does it mean generating an xml file too ? – Jean-noël Lafargue Apr 30 '18 at 08:22
  • (silly question, I just saw that the took I use manges HAAR, LBP and HoG too) – Jean-noël Lafargue Apr 30 '18 at 08:27
  • the HoG cascades are deprecated. Typical way to use use HoG is training an SVM from.HoG descriptors. There are tutorials for OpenCV . – Micka Apr 30 '18 at 11:41
  • There are many tutorials indeed but they don't seem very clear to me on the question of the images to use to feed the software. – Jean-noël Lafargue May 01 '18 at 08:55
  • your images look ok to me, if you have enough different backgrounds. Can you add some sample images of how your object looks like in images where you want to detect it? – Micka May 01 '18 at 09:34

0 Answers0