search

SamsungLabs / rome

Realistic mesh-based avatars. ECCV 2022
Other
431 stars 41 forks source link

About the training dataset #19

Closed xmlyqing00 closed 2 years ago

xmlyqing00 commented 2 years ago

Hi,

Nice work! I'm interested in reimplementing the training process.

  1. Since you use the VoxCeleb2 dataset, do you have any suggestions on downloading the dataset? I found that the official website doesn't support batch-downloading videos.
  2. Do you use VoxCeleb1 or VoxCeleb2 dataset?

Thanks!

khakhulin commented 2 years ago

Hi! We use VoxCeleb2 and we filter dataset with respect to image quality since most head segmentation networks cannot work on noisy images I suggest using the recently proposed https://celebv-hq.github.io/ for the best quality.

xmlyqing00 commented 2 years ago

Thank you! I'll check this dataset. It's really annoying that many youtube videos are missing.

I have the following question when you used DECA to reconstruct the face model, why you not use the detailed map (detail displacement map) from DECA? It should provide more geometry details, right?

Thanks for your time.

khakhulin commented 2 years ago

Yes, you can use this map, but since the map itself is noisy on the face we decided to have a model without it.

xmlyqing00 commented 2 years ago

I see. Thank you for your comments.