Jianzhu Guo
Jianzhu Guo
I use the official model provided by Dlib, not trained it by myself. I recommend using FaceBoxes or other detectors, you can refer to [3DDFA_V2](https://github.com/cleardusk/3DDFA_V2).
You can fit it, e.g., https://github.com/YadiraF/face3d/blob/master/face3d/morphable_model/fit.py.
It seems to be the opencv problem, not the repo. You can check the code logic and find if there exists some dead-loop. Honestly, the graphic system like opencv or...
pdc is easy to overfitting, but your loss gap seems to be larger than mine. You can try to tune the hyperparameters like the learning rate.
The depth image generated may be called `Pseudo-depth`. If you want detailed depth, you may refer to the work for the detailed 3d face reconstruction, or use physical devices to...
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...
Nope so far. But it may not be hard to re-implement it by Keras or TF2.0, and welcome for any contributions.
What does `undamaged archive model` mean?
The default file type of the model trained by PyTorch is `.pth.tar`, there is no need to untar it in PyTorch.
I just say: welcome to try our work.