How to train good cascade?

asked 2018-02-27 05:05:38 -0500

Y2Kot gravatar image

updated 2018-02-27 08:44:16 -0500

berak gravatar image

Hello. I have a task to find a dot on the screen. Can you help me to train the best cascade? What settings should I use? Now maded cascade can't find my dot or chooses a lot of false items on the screen.

Here is my Good and Bad samples:

Example of the image with dot in Good folder:

image description

image description

Example of the image file without dot from Bad folder:

image description

edit retag flag offensive close merge delete


Some comments

  • People will never download your random zip files :D
  • A cascade to detect dots? I am quite sure that is not the best approach.
  • Provide sample images in this topic, so we can help you find the best solution!
StevenPuttemans gravatar imageStevenPuttemans ( 2018-02-27 08:20:06 -0500 )edit

fixed, thank you for your reply. I'm not sure how it correctly name in english - maybe circle will be better.

Y2Kot gravatar imageY2Kot ( 2018-02-27 08:26:05 -0500 )edit

There is no object in your test image?

StevenPuttemans gravatar imageStevenPuttemans ( 2018-02-27 08:29:57 -0500 )edit

first two images is an example of my object, I'll add full Good folder screenshot.

Y2Kot gravatar imageY2Kot ( 2018-02-27 08:32:00 -0500 )edit

i'm afraid, too, that your idea won't work, ever.

your dots seem pretty random, it needs something "rigid", with a well defined texture here (think of a face, or a stop-sign)

also, it can't learn anything from "just black" negative images.

again, most probably it's the wrong tool for your job. try with much simpler tools, like the SimpleBlobDetector, findContours, etc.

berak gravatar imageberak ( 2018-02-27 08:37:33 -0500 )edit

ok, sry. I just tried to explain

Y2Kot gravatar imageY2Kot ( 2018-02-27 08:42:56 -0500 )edit

nvm, it's actually a good image, because it highlights the problem (input data far too random) !

berak gravatar imageberak ( 2018-02-27 08:45:26 -0500 )edit

Good images was made with 2 cameras. Will it work if I'll make to cascade for each camera?

Y2Kot gravatar imageY2Kot ( 2018-02-27 08:49:01 -0500 )edit

sorry, but you'll probably have to look for an alternative method of detecting those dots.

training cascades on this unfortunately seems to be an entire waste of time

berak gravatar imageberak ( 2018-02-27 08:58:35 -0500 )edit

What will be better to use?

Y2Kot gravatar imageY2Kot ( 2018-03-01 01:29:31 -0500 )edit