The images on the right show the
(top 24) most
content
similar images in
the dataset to the generated
image
(on the left).
Similarity is measured by comparing the cosine distance of features extracted (using a
VGG16
network
) from the
generated image and images in the dataset.
Features are extracted from an early layer in the
VGG16
network (block1_pool) and a later
layer (block5_pool) .