Training data

[sh1027]

Thank you for sharing your impressive work!

I noticed you mentioned using the VoxCeleb1 and VoxCeleb2 datasets for training. Could you please clarify:

• How many images were used?
• How did you pair source and target images?

Thanks!

[supersonicMaclaurin]

Have you figure out it ?

[AvirupJU]

I suppose they train the model for self-reenactment under self and cross-reconstruction settings. Randomly sample two frames i1 and i2 and forward(ID(i1), pose(i1)) and forward(ID(i1), pose(i2)). Calculate loss wrt GT i1 and i2.

[diligent1787]

Suppose their released model was trained on the higher quality original videos downloaded straight from youtube. If you use the version of VoxCeleb offered by KAIST which comes in 224x224 .mp4 per video, the recommended video processing library would not output any face crop because of the minimum face size of 256.