Sicheng Xu

Cool, I will check the paper details later and thanks for the reply.

Hi, @xvjiarui , I have read the paper and I am curious about the evaluation for the VIP dataset, **could you share the evaluation code for this dataset?** Something is...

Thanks for you contribution! Before I merge your code to the branch, I have some suggestions about the script. 1) Could you add a flag to handle the debugging section?...

> 您好，我注意到您源码里训练时用的BFM model是经过转化以后只有35709个点的BFM model > > 我根据'BFM_exp_idx.mat'从原始BFM09（有53490个顶点）中提取了53215个点并且与“Exp_Pca.bin"中的表情基进行了整合，最终得到了'bfm_53215.mat'（相对于您用的35709个点的模型，增加了脖子和耳朵，验证过其拓扑结构也没有问题）。 > > 我注意到您在根据回归出来的coeff来重建出个性化人脸模型时，进行了recenter，具体如下（详情可见/model/bfm.py)：  > > 想问若我想使用我得到的'bfm_53215.mat'来重建出含有脖子和耳朵的个性化人脸模型，是否也需要recenter？ > > 若recenter以后，在相同的coeff情况下，得到的含有脖子和耳朵的人脸模型再去掉脖子和耳朵以后 跟 直接将coeff作用于35709个点的BFM model得到的个性化人脸模型，他们的三维点坐标是不一样的。我个人认为的原因是在recenter是求平均值时，因为bfm_53215的meanshape的平均值与bfm_35709的meanshape的平均值不同。 recenter的目的是为将人脸的中心改为(0, 0, 0)，我们训练的网络的pose是基于“bfm_35709”的空间位置预测得到，所以将回归得到的系数直接用于“bfm_53215”，会出现pose不一致的问题，解决办法也较为直接，即利用“ bfm_53215”减去“bfm_35709”的中心即可。

The description is not clear. From the output which is rendered by our code, the left boundary of the reconstructed face is quite aligned with the input. If you adjust...

First of all, I am not familiar with the blender. But I want to point that, we used a common simplified camera model setting. The focal in our projection matrix...

You may try some tutorials of the nvidiffrast to check if it's installed correctly. Or you can follow #12 to replace nvidiffrast with pytorch3d for easier setup.

Could u provide more info about your environment? such as tensorflow version, cuda version, etc.

As shown in the figure blow, the 2nd column is the head region decided by semantic segmentation, including face, ear, and hair. The 3nd column shows the region of rendered...

For your purpose, you can refer to #15 .