Sample Training

Post date: Mar 07, 2012 4:54:32 PM

Now that I've figured out how to do it I found a great tutorial on haar training http://note.sonots.com/SciSoftware/haartraining.html

The point of interest for me is the very nice description of how to use opencv_createsamples to generate a sample set from multiple images. All the rest I've found say 'use createsamples put your images in a directory and create a description file. It should look like this:

img/img1.jpg 1 140 100 45 45 img/img2.jpg 2 100 200 50 50 50 30 25 25 img/img3.jpg 1 0 0 20 20

What do all those numbers mean? I had no fricking idea until about 1:30 am last night.

They're descriptions of bounding boxes which define where in the image to pull the

sample from. But even after learning I couldn't have told you how to create one.

A very nice description in the tutorial above. It's slightly dated so it covers the

command haartraining which has been depricated in favor of traincascade so I'll have

to find or write a good description of how to use that one. Overall the link above

is fantastic

I wish I had found this earlier but unfortunately I needed to learn a whole lot more

about object detection before I knew what to look for. In reality the folks at

OpenCV have made it blitheringly, mind bogglingly simple using their library.

They just haven't told us _how_.