Semantic Facial Image Manipulation using 2D/3D modalities
Sponsored by Fujitsu | Advised by Laszlo A. Jeni
Slides
Mid-Semester-1 Presentation Slides (Sensitive, tend not to publish)
Mid-Semester-2 Presentation Slides (Sensitive, tend not to publish)
Final Presentation Slides (Sensitive, tend not to publish)
Videos
Final Presentation Video (Sensitive, tend not to publish)
Our proposed pipeline mainly consists of two modules, geometric manipulation module and semantic manipulation module. The geometric manipulation module warps the feature maps of input image guided by estimated 3D geometry. In this way, artifacts are reduced by moving the manipulation from image space to feature space. The 3D geometry introduced here helps handle extreme head pose and lightning condition. The warped feature maps are then fed to semantic manipulation module which synthesizes information that was not the input image.
Geometric Manipulation Module
As shown in Figure 1, given an input image and the FACS code of driving image as target expression, we first feed the input image into CNN encoder to extract its feature map. We also estimate the 3D geometry of the input image, which is later sent to the 3D geometry module to output an estimated 3D geometry of target image conditioned on target FACS code. Then we calculated an optical flow between input image and target image according to their estimated 3D geometry. Finally this optical flow is used to warp the feature map of input image.
Semantic Manipulation Module
Besides the estimated 3D geometry of target image, the 3D geometry module also outputs an inpainting mask which synthesizes information that was not the input image and locates areas that need to be inpainted (e.g. wrinkles and teeth). The Semantic Manipulation Module first masks the warped feature map with inpainting mask and then fuses this masked feature map with target FACS code. After fusing, the feature map is fed to an CNN decoder to get the generated image.
Supervision
Our final loss function is the combination of perceptual loss, geometry loss and GAN loss as shown below. I denotes image, P denotes 3D geometry, D denoted discriminator.
Qualitative illustration
Individual AU Editing result
Qualitative comparison with GANimation
Semantic Facial Image Manipulation is a conditional generation task. The goal is to synthesize a facial image conditioned on both identity and target expression (In our settings, identity information is provided as a 2D image and target expression is provided as a FACS code).
Social Motivation
Fujitsu is interested in using FACS-based expression analysis to analyze viewer’s behavior for advertising and to provide feedbacks for video conferences, remote training, etc. But FACS annotated data is limited in amount and often suffer from skewed distribution. Some expressions and Action Units are much more frequent than others. Using facial expression manipulation can generate FACS annotated data while controlling the distribution of generated data (i.e. we can generate expressions and Action Units we want).
Technical Motivation
2D based methods like GANimation [1] manipulates faces in 2D image space, which often results in artifacts in generated images and failures under extreme head pose and lighting condition.
Proposed Pipeline
We proposed an encode-manipulate-decode pipeline mainly consisting of two modules, geometric manipulation module and semantic manipulation module. The geometric manipulation module warps the feature maps of input image guided by estimated 3D geometry. In this way, artifacts are reduced by moving the manipulation from image space to feature space. The 3D geometry introduced here helps handle extreme head pose and lightning condition. The warped feature maps are then fed to semantic manipulation module which synthesizes information that was not the input image
References
[1]Pumarola et al. GANimation: Anatomically-aware Facial Animation from a Single Image, ECCV2018
[2]Ekman, Paul, Wallace V. Friesen, and Joseph C. Hager. “Facial action coding system: The manual on CD ROM.” A Human Face, Salt Lake City (2002): 77-254
Zhuoqian is an MSCV student of class Fall 2019. He received his bachelor’s degree in Software Engineering from Beihang University. His research interest is in creative computer vision, content creation and manipulation with generative models and unsupervised/self-supervised learning.
Responsibilities: Data processing, pipeline construction, experimentations and analysis
Dai is an MSCV student of class Fall 2019. She received her bachelor’s degree in Department of Automation from Tsinghua University. Her research focuses on object detection, semantic segmentation and network interpretability.
Responsibilities: Experimentations, qualitative & quantitative evaluation and analysis
Advisor: Laszlo Jeni
Laszlo is a Systems Scientist (faculty) in the Robotics Institute at Carnegie-Mellon University. I focus on advancing the state-of-the-art in multi-modal methods for computational behavior science, specifically in areas of modelling, analysis, and synthesis of human behavior and emotion using diverse sensors.