In this framework, we provide an implementations for the class ifeature
that are able to read the features produces by two CNN, namely OverFeat (paper) and VGG-16.
As a pre-processing step, we extract the features from the nets and store them into the '*.txt' files.
The following classes know how to read and binarize certain features:
-
cnn_feature.h
reads the OverFeat features and perfoms mid_binarization -
cnn_feature_mean.h
reads the OverFeat features and perfoms mean_binarization -
vgg_feature.h
reads the VGG-16 features and perfoms mid_binarization -
vgg_feature_mean.h
reads the VGG-16 features and perfoms mean_binarization
For details on binarization, check relocalizers.
For details on '.txt' formats check examples.
To use your own features, you need to derive a class from ifeature.h
. If you want to use relocalizers you should derive from ibinarizable_feature.h
.
- You should reimplement all
virtual
functions. - You should update the
feature_factory.h
to let it know, which type of features to work with. Within the main framework all the feature generations are done throughfeature_factory.h
.