juyong / 3dface Goto Github PK

Hi juyong! I read your paper and wrote a small script to fit id and expression parameters. However I have several questions.
1.In equation 5, if I fit only using landmark term, I can fit at most 40-d id param and 10-d exp param. If longer, the algorithm does not converge.
2.Does the photo consistency term help for the precision of id param and exp param? Or they are only used for the albedo param.
3.How to use Gauss Newton to minimize a function with form like: f(x)=1/MSigma(||g(x)||^2)+1/NSigma(||h(x)||^2)?
Thank you very much!

Can you provide a direct download link for CoarseData that I can use through wget? I need this exact file: Exp_Pca.bin

In the dataset you given to us, we can get the 3DMM parameters of coarse dataset and the depth of fine dataset. As we can see in the paper, the 3DMM parameters of coarse dataset is derived from the first picture of fine dataset, which is the original picture.What I want to get is the 3DMM parameters of the original picture which is not rendered by the transformed 3DMM parameters.I want to use my code to get the depth of the original picture and compare with the corresponding depth you given to us in the dataset to check my code is right or wrong.Thanks for your kindness consideration.

We can see that the 3DMM parameters you given to us in your dataset is very accurate.I find some open source code in the website and I want to recover the parameters of any face, but all of them can not get accurate parameters as you.Could you give us the source code of geting the 3DMM parameters of faces?Thanks a lot!

Hello, I wanna know if ExpBasis has closed eyes exp？

Hi, I am wondering how is the last 6 parameters ordered for the 185 parameters? I suppose it should be pitch, yaw, roll, t_x, t_y and s. In that case, how could we convert the float number of pitch, yaw and roll to the degrees? Thanks!

Hi, I have down load Exp_Pca.bin, and found that the vertex number is much larger than FaceWarehouse, in which each face mesh has 11K vertices. As the expression principle axes is computed by using FaceWarehouse, why the vertex number is inconsistent? Thank you.

In the paper, the size of input image is (224, 224).
But, the image' size in Coarse dataset is (450, 450).

Do you preprocess the image before model is trained ?
Crop? Resize ?

Could you tell me about that ?

Thanks a lot!

Does the FineData have the Coarse-like groundtruth file? I want to create Coarse depth by myself.

I read the paper and might need your data in my project. So when will that be online? Thank you very much and that is a good paper.

Hi, your work is great!　I've downloaded your Exp_Pca.bin and try to use it to generate expressions. I wonder why the point number is 53215 while the original BFM model has 53490 vertex. Could you show which vertex do you preserve. Besides, can you include the Eigen value of expression bases in the model?

How many expression basis contains inside it? and how does they are defined?

can expression express eye blink? i find https://github.com/microsoft/Deep3DFaceReconstruction which use your expression can not express eye blink

Hi, for the training data of the Tracking CoarseNet, it seems to be generated randomly. I am wondering could you also provide the parameter of the normal distribution you used to generate the training set? Thanks!

I find the .obj file in results only has 31513 points. Can you provide the index corresponding to BFM (which has 53490 points). Thank you very much!

In the paper,we can find how to calculate the 3D face geometry by p = p¯ + Aidαid + Aexpαexp.But in the dataset you give to us we can only find 100 αid parameters, we cann't find the αexp to reconstruct the albedo of the model.Is there anything wrong when I use the dataset or the parameters is inconsistency?

Thanks for the sharing of your great work. I wanted to replicate your experiments, but it seems that the links for the datasets that you used in the paper and for the BFM on your github description are expired. Could you help out?

Hi, I am interested in what is the range for the loss of single coarse net? The ground truth parameter seems to be huge compared to traditional deep learning problems.

"The requested URL was not found."
https://drive.google.com/open?id=1z12Ukg-iIF4qM3z6RXKUzYYpl43seO2J

It seems only datasets for training Single-image CoarseNet and for FineNet are included in this repo.
Is the dataset for training Tracking CoarseNet (extended from Single-image CoarseNet data by simulating previous frames and also including albedo and lighting parameters) included?

Hi!
Thanks for sharing your dataset and results to public.
I really appreciate it.
However, I could not download the files with the provided Baidu link.
I was unable to make an account from my country,
and the link itself would not let me download any files without an account.
Could you please provide me with an google drive link?

Thanks in advance.