Facial detection positive sample vector size
After looking at the documentation, papers and tutorials for creating a LBP cascade, I still don't know how the image size of vector file created from positive samples affect the cascade created. Here's the command I'm talking about if it's not clear, and specifically width(-w) and height(-h) values.
opencv_createsamples.exe -info p_samples.info -num 1000 -w 20 -h 20 -vec ps.vec
I've two questions about this.
First one, does cascade detect faces smaller than width and height values defined?
And second one, does increasing the size of the width and height values increase the quality of the cascade created? If yes, is there a point of diminishing returns or one should use largest possible size if training time is not an issue?
First one: no.
larger size means more information/features, so it may have more noise, too. I think it is a matter of testing. Try 3-4 sizes (10x10 15x15 20x20 30x30) and see what are the results. But there is no way you can detect the 10x10 face with the 20x20 cascade, except you are resizing the initial frame/photo 2 times