What is trainingfaces_24 -24.vec? [closed]

asked 2014-01-28 12:34:30 -0600

xaffeine gravatar image

In the data/vec_files directory, there is a file called trainingfaces_24-24.vec. Is this the file used for training the face-dectection files in haarcascades, or just an example file to be used for similar purposes?

edit retag flag offensive reopen merge delete

Closed for the following reason the question is answered, right answer was accepted by sturkmen
close date 2020-10-26 17:37:28.777730

Comments

It is the actual data vector used to train the internal frontal face haar wavelet cascade that comes with openCV. It is there for people to try and get similar results.

StevenPuttemans gravatar imageStevenPuttemans ( 2014-01-29 08:26:17 -0600 )edit

Thanks, that is good to know. Does it correspond to haarcascade_frontalface_default.xml, haarcascade_frontalface_alt.xml, lbpcascade_frontalface.xml, or somehow all of the above? Also, does the data come from a generally-accessible database or is it a custom collection from various collections?

xaffeine gravatar imagexaffeine ( 2014-01-29 12:44:04 -0600 )edit

As far as I know it is used for training the haarcascade_frontalface_default and the lbpcascade_frontalface. No idea where they got the data, but there are tons of sets out there to get similar results if you want to test it. Also, using bootstrapping with hard negatives can help you to get far better detectors, with less false positives.

StevenPuttemans gravatar imageStevenPuttemans ( 2014-01-30 04:27:39 -0600 )edit

Did you mean to say that I could get detectors far better than haarcascade_frontalface_default? I would like that very much!

xaffeine gravatar imagexaffeine ( 2014-01-30 12:40:23 -0600 )edit

Yes you can. Start by grabbing a large dataset, set it up correctly and apply like 5 stages of negative bootstrapping. You will notice the difference immediatly. Its what commercial compagnies do for their cameras.

StevenPuttemans gravatar imageStevenPuttemans ( 2014-01-31 01:54:07 -0600 )edit

Cool. By "large" do you mean the size of Feret? Or is bioid+muct big enough?

xaffeine gravatar imagexaffeine ( 2014-01-31 12:39:24 -0600 )edit