Ideal Image Parameters and Features for Object Detection

asked 2016-02-29 13:14:21 -0500

jallen gravatar image

I'm building a cascade to detect electrical outlets. I've gone through the 3 or so available tutorials, which are helpful.

What I haven't found and am still unclear on is what are the ideal traits of the positive and negative images that would result in the most perfect object detection?

I have numerous questions like:

  1. Does the object need to comprise ALL of the positive image?
  2. Can there be variations of the object in the positive image ( I assume so because for faces we want to supply many different faces) ? For instance, provide a positive image with an outlet that has one outlet plugged in?
  3. Sizing: should I presize all of the images? Meaning should all of the positive and negative images be the same size?
  4. what about the distortion options opencv_createsamples provides? I rotated 180º but what about coloring? How do I want to recolor my images to achieve the best results?
  5. Regarding this dataset of negative images here: should I size my positive images to the size of these negative images? opencv_traincascade gave me numerous errors when training with these negative images, possibly because they were different sizes.

I think you can see my line of questioning, and I realized there are multiple questions here. It's just I haven't found information about what I should ideally provide to OpenCV for training. I think that I'm actually seeking better docs, because the docs don't really provide more information about the positive and negative images.

One more thing: I think because of the time it takes to train here, a trial-and-error approach to improving the training is not optimal. Plus there are many different variables as far as training that it would take weeks to optimize this on one's own.

edit retag flag offensive close merge delete