I will train haar cascade to be able to detect an object that works fast. I prepared positive images to be used for training. The question I want to ask is about how I should choose negative images. Should negative images be another object that is not the object to be detected, or just a background? While processing negative images, does it scan the negative image like when detecting, or does it consider the negative image as an object and resize it to the window size and apply the filters accordingly. As a result, should I separate the negative image background image into subsamples and train them as separate images or should I train the background image as it is without breaking it into pieces? What do you suggest?
To make my question more descriptive, I tried to explain it with image below.
forget that idea, it will never work.
cascades only work with rigid, (consistently) textured things, – your ping pong paddles are not like that.
yea, more lke that
no, it will do that internally
but again, you picked a very hard problem.
given the real life pose variation of those paddles, even cnns will have a hard time
I know the difficulties of Haar cascade and I’m trying to clear things up to improve accuracy.
The table tennis racket was an object that I showed as an example to explain my question.
As I understand from what you said, in training, it divides the background image into grids and treats each one as an object. That is, it does not perceive the entire background image as an object by shrinking it, instead it divides the image into grids and processes each grid as an object. Do I understand correctly?
exactly like that …
(so you want large neg images here)
Thank you for your answers.