0

And again about cascade training. I have set of images with resolution approximately 1000x1000. Object size is nearest 30x50. What is better for background images: take it as is or crop to smalles like 100x100?

PS. I use opencv train_cascade utulily. Images are satellite imagery. Now use Haar-like features.

Bleach
  • 156
  • 3
  • 15
  • background of positive samples or negatives? For negatives use full size (more subsamplinfs possible), for positives it should make no difference, since only the annotated object area is used. – Micka Jun 27 '18 at 15:08
  • @Micka. Thx. I meam what background is a negative samples. From positives taked only croped and resized to mask size object to find. – Bleach Jun 27 '18 at 16:51

0 Answers0