search

cleardusk / 3DDFA

The PyTorch improved version of TPAMI 2017 paper: Face Alignment in Full Pose Range: A 3D Total Solution.
MIT License
3.6k stars 651 forks source link

How to generate the ground-truth pose parameter? #163

Open zengxianyu opened 4 years ago

zengxianyu commented 4 years ago

Thank you for the amazing work. I am very interested in it and would like to ask some questions about the details.

From the training code it seems that the model is regressing the pose, shape and expression parameters. By my understanding, the shape and expression ground-truth parameters are from the bilinear face model, but how to get the ground-truth pose parameters?

Another question, is there any way to derive the expression blendshapes or blendshape weight scores from the model output given an arbitrary input face image?

I am looking forward to your reply and thank you in advance.

zengxianyu commented 4 years ago

Are the ground-truth pose parameters generated by data augmentation using face profiling?

cleardusk commented 4 years ago

The gt data is fitted by analysis-by-synthesis, including the pose, shape and expression. Faceprofiling is used to augment the dataset, the parameters are also generated correspondingly. For blendshape, you may need to build by yourself or find the pre-defined ones, this repo does not define the blendshapes.