We propose DeepExpr, a novel expression transfer approach
from humans to multiple stylized characters. We first train two Convolutional Neural Networks to recognize the expression of humans and
stylized characters independently. Then we utilize a transfer learning
technique to learn the mapping from humans to characters to create
a shared embedding feature space. This embedding also allows human
expression-based image retrieval and character expression-based image
retrieval. We use our perceptual model to retrieve character expressions
corresponding to humans. We evaluate our method on a set of retrieval
tasks on our collected stylized character dataset of expressions. We also
show that the ranking order predicted by the proposed features is highly
correlated with the ranking order provided by a facial expression expert
and Mechanical Turk experiments.