Haar Training: error (-215)_img.row * _img.cols == vecSize in function get

edit

Thank you for responding. I ended up actually figuring out this error. It was the fact that sampleWidth and sampleHeight were 24 ( as can be seem in the error statement above), but I was giving it 80 and 40. Which makes it seem like 24x24 is some sort of vector size default? However, I would still like to know how I can improve my training performance. To answer your question I am using a famous perl script called createsamples.pl by Naotoshi Seo.This mixes the positive and negatives and does apply warping. However, for some reason no matter how many images I have, it only creates 1000 vecs. This is the only way I know how to create vector files. Is there an alternative to this? What is a better minHitRate? These are all parameters from a tutorial.Thank you again.

This is my latest command of how I created my vectors:

perl bin/createsamples.pl positives.txt negatives.txt samples/ 7000 "opencv_createsamples -bgcolor 0 -bgthresh 0 -maxxangle 1.1\ -maxyangle 1.1 maxzangle 0.5 -maxidev 40 -w 24 -h 24"

Although it only creates 1000vecs even though I specified 7000, and might have some unnecessary slashes.

AHA I know your problem. -w and -h parameters are defined at the time you are creating your sample set. If you set then 24x24 there, then do not expect your training to work with 80x40. Those dimensions need to match exactly since they define how your training setup runs through the data vector. Also, that script is the worst script ever, it makes absolutely NO sense to combine positive and negative training samples into a single training vector ... no idea where that dude got the inspiration ... the 1000 is a parameter in the createsamples interface. Use the C++ version from the OpenCV repo and not that perl script.

Thank you for your help. Strangely, even when I used -w 80 -h 40 for the pearl script and used -w 80 -h 40 on traincascade I still got the error. Using 24 x 24 works now, but it seemed as if 24x 24 was some sort of default? I don't know if matters much since I won't be using that pearl script again, but it seemed as if traincascade required 24x24 vec samples. So for using the C++ version I've read that I have to create .dat file with the location of the image of interest next to the name of the pic. Is there any fast way to get the location of the image of interest from every single positive pic, or should I just choose pics with similar locations of the object of interest and use the same description for every positive?

Probably the perl script doesnt overwrite the parameter and simply takes default value. As for the locations, you will need to use an annotation tool to store and record all locations. Object detection is a hell of a work :P

ok, thank you very much

I have another question if you don't mind. When I tried doing

opencv_createsamples -info positives.dat -vec samples.vec -num 2387 -w 24 -h 24

I had 2387 positive images, I know I probably need more, but the issue is I still only got 1000 samples in the vec file

   Info file name: (NULL)
Img file name: (NULL)
Vec file name: samples.vec
BG  file name: (NULL)
Num: 1000
BG color: 0
BG threshold: 80
Invert: FALSE
Max intensity deviation: 40
Max x angle: 1.1
Max y angle: 1.1
Max z angle: 0.5
Show samples: FALSE
Width: 24
Height: 24

Are you sure that positives.dat lists more than 1000 items?

When I scroll to the bottom it says line 2387, which is how many positives I have .