caulier.gilles at gmail.com
Wed Jul 16 19:52:04 BST 2014
2014-07-16 20:06 GMT+02:00 Marcel Wiesweg <marcel.wiesweg at gmx.de>:
>> This string returned by this method is wrong i think as TLD algorithm
>> is fully replaced by LBPH:
>> Right ?
> Yes, it's wrong
Fixed : http://commits.kde.org/libkface/bbe822aeca1a2267a67cda283ebe347d78eb6010
>> >> For some reason, the necessary dummy updater methods
>> >> (updateV1ToV2) are already present.
>> > I can remove it if necessary, but this method is called by makeUpdates()
> I meant to say, we need these mehods now, they were there although not needed
> previously. All is well.
>> Do you take a look to OpenBR project in the past ?
> No, I have not seen this project yet..
> After a quick read through the paper they mention, they use Viola-Jones faces
> detection (as we do), normalized by the detected eye location (brings
> difficulties in my experience, as the eye is often not detected) and by Tan-
> Triggs (simple, implemented in libkface). Recognition is based on LBP (as we
> do) and, in addition, SIFT descriptors and some improvements in the matching
Very interesting. The technic explain on OpenBR site is very
instructive (i never look how form recognition and detection work
If i well read the code, there is a part, always compiled in libkface
but not used due to bad results : funelreal.cpp
It's to Align, Normalize, and Flat face elements for better analysis
in case of train and later recognize using LBPH.
If it's something like that, this part must be improved in the future,
through GSoC. I'm sure that we can find a student easily to work on
face management algorithm.
Or perhaps you have others ideas to improve libkface in the future ?
More information about the Digikam-devel