train object detector for face detection

[Sagar Patel]

Hi,
I want to detect faces from images.For that i am using dlib c++ library.
For that i am creating my own object detector using train_object_detector.cpp and fhog_object_detector.cpp.Suppose i am training detector using 10 sample images,so after completion of training it will create one object_detector.svm file. Now again if i train detector using another 10 images then it will again create new svm file.But i don't want that.I want to append new data to existing svm file.
So, my question is , it is possible to append new trained data to existing svm file.In dlib pegasos is there.I have googled and found that pegasos do this.But i can't figure it out.Can you please give me some hint for that.

[Davis]

It doesn't work that way, even if you used pegasos.

If you have new training data you need to run training again with all the
data to get a new .svm file. The old .svm file is not involved.

[Sagar Patel]

It is possible with any other library?

[Davis]

No, what you are asking mathematically just doesn't work out.

[Sagar Patel]

So everytime when i train new data ,previously generated .svm file will be overwritten?

[Davis]

Yes

[Sagar Patel]

ok.Thanks davis.

[Sagar Patel]

and what about online and incremental algorithm?
and another question is: I am trying to train object detector with around 3000 input images, the problem I have is my RAM 4GB gets overloaded and the training program gets crashed. So, I started finding ways by which I can train images in batch and create combined svm. Is there any alternative to memory issue while training with large amount of data?

[DomLi]

Hi Davis,

I am always curious about your wonderful job, I just wanna ask if there is any GPU (cuda) support on the training SVM model with fhog_object_detector_ex.cpp now?

[Davis]

No, that tool is cpu only.