0

I train the cascade classifier to detect pedestrian traffic lights, I used 5000 positive photos and 1000 negative for training, here is the result xml file I got.

The problem is the trained cascade detects many false images. I used the original code of OpenCV face detection, and I also notice that the best result is when I set minneighbors in detectMltiScale to 100.

How should I increase the trained cascade quality?

Danaro
  • 61
  • 1
  • 2
  • 11
  • use (much) more negatives. Typical negative-counts are 2x number of positives IN EACH STAGE! If possible use negatives that represent your typical backgrounds. If possible, post some of your positives and negatives here to give an impression of whether the data is suitable. – Micka Jun 29 '16 at 10:07
  • What is the ultimate image size for running detection on? – Danaro Jun 30 '16 at 06:31
  • I would say that the optimal image size is the image size where your target object is slightly bigger than the training size. But that's just my intuition – Micka Jun 30 '16 at 06:35

0 Answers0