240 likes | 383 Views
1. Facial Expression Editing in Video Using a Temporally-Smooth Factorization. 2. Face Swapping: Automatically Replacing Faces in Photographs. Facial Expression Editing in Video Using a Temporally-Smooth Factorization.
E N D
1. Facial Expression Editing in Video Using a Temporally-Smooth Factorization 2. Face Swapping: Automatically Replacing Faces in Photographs
Facial Expression Editing in Video Using a Temporally-Smooth Factorization FeiYang, LubomirBourdev, Eli Shechtman, JueWang, DimitrisMetaxas CVPR 2012
Goal The goal is to allow for semantic-level editing of expressions in a video: • magnifying an expression • suppressing an expression • replacing by another expressions
Challenges • Natural expression • Different parts changes accordingly • Unique identity • Temporal coherency
Related Work • 2D based methods • [Theobald09], [Liu01], [Williams90], … • 3D based methods • [Blanz03], [Pighin98], … • Expression flow • [Yang11]… • Frame reorder method • [Bregler98], [Kemelmacher- Shlizerman11] • Tensor factorization methods • [Vlasic05], [Dale11]…
Algorithm 3D Tensor Model - [Vlasic et al siggraph05] Modify Expression Information Identity Information
Algorithm goal to identify a and 2D v.s. 3D method = Weak Projective Matrix Rt frame t Minimize: | – |
Algorithm Fitting Error: Shape Distribution Constraint: Temporal coherence:
Algorithm Levenberg-Marquardt (Siggraph98)
Algorithm Adjust to achieve expression modification • Dynamic Time Warping (DTW) [Sakoe78] • Residual Expression Flow • Correcting boundary compatibility
Face Swapping: Automatically Replacing Faces in Photographs Dmitri BitoukNeeraj Kumar SamreenDhillon Peter Belhumeur Shree K. Nayar Siggraph 2008
Goals For an input image: • Automatically find the best candidate • Automatically replace the face • Automatically color and lighting adjustmet
Library Building OKAO face detector to detect face pose [Omron07]
Alignment Pose, Resolution, and Image Blur: • Yaw, pitch threshold between two images ( ) • Eye distance as a measure of distance (80%) • Similarity of the blur degrees [Kundur and Hatzinakos 1996; Fergus et al. 2006]
Color and Lighting To ensure the similarity between the replaced and original face, a linear combination of 9 spherical harmonics [Ramamoorthi and Hanrahan 2001; Basri and Jacobs 2003] is used as measure metric: Each pixel I(x, y) can be approximated by: Distance:
Seam Signature 256-by-256 patch from the face is used for replacement. Unfold: L2 Norm is used to compute the distance
Appearance Adjustment Using simple scaling on the Harmonics coefficients , are the original and replacement images Scale the replaced image
The End Any Questions ?