search

tencent-ailab / hifi3dface

Code and data for our paper "High-Fidelity 3D Digital Human Creation from RGB-D Selfies".
Other
756 stars 153 forks source link

Non-Asian .mat files #17

Open razzraziel opened 3 years ago

razzraziel commented 3 years ago

So your AI-NExT-Shape-NoAug.mat, AI-NExT-Shape.mat etc constructed/trained from 200 asian models. Is it possible us to create non-asian versions? Or both for global use?

cyj907 commented 3 years ago

So your AI-NExT-Shape-NoAug.mat, AI-NExT-Shape.mat etc constructed/trained from 200 asian models. Is it possible us to create non-asian versions? Or both for global use?

You can use other 3DMM model, such as BFM for constructing people of other races than Asian.

razzraziel commented 3 years ago

Hi @cyj907

I started to test this project again. Achieved results were exactly same with your output. Great job.

But unfortunately, as you already mentioned in your paper, non-asian faces result with lack of face details.

Is there a way to train my own network similar to Asian Mats? Any training workflow integrated in this project? If not, can you show me a way to follow, so I can use photos to create my own data similar to yours?

cyj907 commented 3 years ago

Hi @cyj907

I started to test this project again. Achieved results were exactly same with your output. Great job.

But unfortunately, as you already mentioned in your paper, non-asian faces result with lack of face details.

Is there a way to train my own network similar to Asian Mats? Any training workflow integrated in this project? If not, can you show me a way to follow, so I can use photos to create my own data similar to yours?

Sorry for the late reply. Our pipeline is not based on training neural networks, but based on direct optimization of paramters. You can use other 3DMM models like FLAME or BFM to train for people of other races. If you capture depth data, you can directly use our model to reconstruct Westerners and Africans, etc. We have tested already, but do not show the results here.

hrx000 commented 1 year ago

So your AI-NExT-Shape-NoAug.mat, AI-NExT-Shape.mat etc constructed/trained from 200 asian models. Is it possible us to create non-asian versions? Or both for global use?

Hii, have you trained a model for non -Asian entities?