Our main contribution is the idea to learn
relative visual attributes, which to our knowledge has not
been explored in any prior work. Our other contribution
is to devise and demonstrate two new tasks well-served by
relative attributes: (1) zero-shot learning from relative comparisons,
and (2) image description in reference to example
images or categories. We demonstrate the approach for both
tasks using the Outdoor Scenes dataset [11] and a subset of
the Public Figure Face Database [12]. We find that relative
attributes yield significantly better zero-shot learning
accuracy when compared to their binary counterparts. In
addition, we conduct human subject studies to evaluate the
informativeness of the automatically generated image descriptions,
and find that relative attributes are clearly more
powerful than existing binary attributes in uniquely identifying
an image.