2

I have problem with creating classifier for handwritten number recognition. I am stuck on first step, creating *.vec file with positive samples. I am working with OpenCV library, version 3.

I am using opencv_createsamples tool, I have dataset with 60k positive samples from MNIST. I created .info file with paths to images and with correct information about rectangle and number of matched object

Format of info is:

    ...
    img/1.jpg 1 0 0 28 28 
    img/2.jpg 1 0 0 28 28 
    ...

And calling tool with

opencv_createsamples -info positives.info -vec digits.vec -num 60000 -w 24 -h 24

Digits.vec file is created without any error but when I try to see what is in file with 

opencv_createsamples -vec digits.vec -w 24 -h 24

It shows me only one sample.

And afterward when I try to run classifier training it show error that there is not enough samples in .vec file.

I tried everything and have no idea what to do. Tried with perl script which creates .vec file for every provided positive image. But there is necessary mergevec tool for merging all those .vec files, which I cant compile because in v3 of OpenCV there is no .cpp files necessary for compiling.

I don't know what to do, can somebody help me? Thanks,

EDIT:

I solved problem with downgrading to 2.4.9 version of OpenCV library. Same process but results are correct.

Steva
  • 231
  • 4
  • 16
  • What operating system / development environment are you using? It's possible this could be involved with '\r\n' vs. '\n'. – ross Jul 03 '15 at 16:28
  • I am using kubuntu 15.04 64bit and every file is created on kubuntu. – Steva Jul 03 '15 at 17:10

0 Answers0