Personalized Facial Action Transfer (FAT)

Transferring facial actions from a source subject to the target from one target image with netural expression. [ECCV, 2012]

Simple Idea:

Directly copy the appearance changes of the source to the target subject, produce artifacts on the target ... Not Working

Key Idea: personalize the appearance changes for FAT

1, Learn a regression to predict appearance changes from shape changes (with respect to the neutral);

2, Personalize the regression using one target neutral face image.

Two Steps:

Step(i): Transfer shape changes from the source by jointly applying the local triangle deformation to the target neutral face ;

Step(ii): Estimate the appearence changes from the shape changes using a personalized regression for the target person.


Source Exp: face of the source subject with expression

Target Neu: face of the Target subject with netural expression

Copy: (Simple idear abve)

Per-Spec: FAT by a regression learned from samples of the target subject (difficult to get the data; may fail to represent the facial action transfered from the source)

Generic: FAT by a regression learned all data of some training subjects (represent the averaged appearance changes of the training subjects)

Personal (Our approach): FAT by a regression persalized only from one neutral face of the target subject

I, FAT between different subjects:

Source Exp Target Neu Copy Per-Spec Generic Personal
FAT Videos: Video 1, Video 2, Video 3, Video 4

II, Facial De-identification: removing the facial feature of the source subject while preserving the facial action

Source Exp Target Neu Per-Spec Generic Personal
De-identification Videos: Video 1, Video 2, Video 3, Video 4