Image preprocessing for inference on VoxCeleb dataset

[StelaBou]

Hello,

I am trying to evaluate your model on VoxCeleb dataset, however the results are poor. I have preprocessed the dataset using https://github.com/AliaksandrSiarohin/video-preprocessing and I run your inference script using --if_extract and --if-align arguments.

Is something wrong with the preprocessing of the facial images? Additionally, is your model able to handle the roll angle of the head pose?

Thank you!

[FeiiYin]

Since our model is finally trained on the HDTF dataset in an end-to-end way, the performance is quite related to the distribution of the dataset. The roll angle in HDTF varies little, hence our model may get poor results when there exists a large pose changement in the driven video. The problem may be solved via training on datasets with larger pose distributions.