Just as an overview, you wouldn't really train overfeat itself on your new images. What Overfeat can do however is instead of make the final "classification" of the input images by saying which classes (that overfeat knows) it belongs to, instead you take the final feature layer.
So you will train your own model, on your own images, using OverFeat features. You can think of OverFeat as returning a massively compressed version of your images, so now you only have a few parameters to train on, instead of raw RGB images.
As to how exactly to do that, it is in the documentation.
So while the final feature layer of OverFeat might "say", this image
-has a head
-has an eye
-has grass
-is brown
-SO IT MUST BE THE GRASSHOPPER (=Final Classification)
you will train you model so that when overfeat gives the features
-has a head
-has an eye
-has grass
-is brown
-IT MUST BE MY DOG REX (=Final Classification)
Overfeat will find good features for you to train on,
Just as an overview, you wouldn't really train overfeat itself on your new images. What Overfeat can do however is instead of make the final "classification" of the input images by saying which classes (that overfeat knows) it belongs to, instead you take the final feature layer.
So you will train your own model, on your own images, using OverFeat features. You can think of OverFeat as returning a massively compressed version of your images, so now you only have a few parameters to train on, instead of raw RGB images.
As to how exactly to do that, it is in the documentation.
So while the final feature layer of OverFeat might "say", this image -has a head -has an eye -has grass -is brown -SO IT MUST BE THE GRASSHOPPER (=Final Classification)
you will train you model so that when overfeat gives the features -has a head -has an eye -has grass -is brown -IT MUST BE MY DOG REX (=Final Classification)
Overfeat will find good features for you to train on,