Bartis Áron's profile - activity

Have you tried reaching the pixels with this method? http://docs.opencv.org/doc/tutorials/core/how_to_scan_images/how_to_scan_images.html#the-efficient-way

My question is that when I use traincascade, there are these two parameters: -precalcValBufSize and -precalcIdxBufSize.

Can you give me some hints about how to calibrate these numbers? I know that the more memory equals lesser execution time, but how can I set a good balance between the two values when the memory is limited?

Also: how much is too much?

Two more questions: what is the size of your negatives and how much free memory do you have? I personally met with this kind of problem and the solution was that I had to close some of my applications (in spite of my 4 GB total memory).

Ps.: Sometimes the traincascade can consume at least 1 GB.

I tried to train with traincascade until the 8th stage. I know, it is few. Now that's why I want to reteach the classifier until 12th stage.

My first question is that if I do not modify other parameters than the maximum number of stages, will the first 8 stage change? I compared the output results, but at the time I did not noticed any difference (except precalculation time).

If no, my second question is that, can I reteach only continuing from the 9th (number 8) stage to get a working cascade? And if I can reteach, how? I could not find any parameter related to this.

I know, there was more post related to this issue, but I couldn't get clear answer to my question.

Thanks!

I must agree that my method was not one of the easiest, but why is a bad choice anyway?

Yes, I meant that, but I thought, that I can solve it automatically either with traincascade or createsamples.

I've searched, but couldn't find any reference. Has the newer "traincascade" any concern about the symmetric images? For example logos can't be mirrored, but faces yes and thus the detection could be improved. I know that about the "haartraining" that it has any, but the "traincascade"?

What kind of error?

You should try the method backprojection: http://docs.opencv.org/doc/tutorials/imgproc/histograms/back_projection/back_projection.html It will give you a probability map of your image, where you get skin color, and where not.

Maybe you should check the size or number of your negatives. I'm not really sure about the haartraining (it is deprecated and the traincascade is more advanced), but if it works like the traincascade, than the number of the negatives doesn't mean too much. Suppose only one big negative (say 480 * 110) then it contains one hundred times more data than one of your positives. That could slow greatly the procedure.