convolutional neural network (CNN) features have been successfully employed
in recent works as an image descriptor for various vision tasks. But the
inability of the deep CNN features to exhibit invariance to geometric
transformations and object compositions poses a great challenge for