Giters

jiaxiangshang / MGCNet

Self-Supervised Monocular 3D Face Reconstruction by Occlusion-Aware Multi-view Geometry Consistency[ECCV 2020]

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Questions about pose of dataset

yjhong89 opened this issue · comments

commented

Hi. Thanks for great work!

I have few questsions of dataset.

  1. Is there any other dataset for training rather than 'Multi-Pie' and '300W-LP'?
  2. How get ground-truth pose data (camera extrinsic, rotation/translation) for '300W-LP' datasets? Isn't it wild images ?
  • I have not downloaded 'Multi-Pie' dataset, but I think this dataset might have gt pose data

Thank you.

Best regards,
YJHong.

commented

For Q1: Facescape dataset, multi-view dataset is expensive and indoor setting
For Q2: We do not use the gt pose, we predict pose and leveraging multi-view geometry consistency.
I do not think so, they have not

commented

@jiaxiangshang
Thanks for replying.

For Q2, then how predict pose vector (rotation, translation) ? Is there any code reference that you would recommend ?

commented

the mgcnet predict the pose actually, you can also see DECA, Deep3DFace......

commented

@jiaxiangshang
Thank you.

These are things that I understand so far.

  • We need rotation/translation/depth information to leverage multiview-geometry consistency.
  • Since ground-truth rotation/translation/depth information doesn't exist for 300W-LP, and MPIE, estimated pose vectors from 3dmm module (DECA, MGCNET, Deep3DFace) are used for multiview-geometry ? (below Equation)
    image
commented

Yes, that is the main idea of MGCNet... :)

commented

@jiaxiangshang Thank you! It is very clear!