Input to SyncNet

[DomhnallBoyle]

Hi, I'm a bit confused about different implementations.

The SyncNet paper suggests that the image input to the SyncNet model is of mouth images and you describe the procedure to extract the mouth region here in this issue: #1 (comment)

Although, in the official repo the author mentions that the input to the pre-trained model is actually the full face and not the mouth region. They don't implement the cropping of the mouth region in their code: joonson/syncnet_python#9

So is the pretrained SyncNet v4/v7 models for frontal and multi-view on the website here actually on the entire face? Wouldn't your code give inaccurate results for the model outputs since you use the mouth region instead of the full face?

Thanks

[voletiv]

Hi,

I believe joonsoon's code was released after the paper, so perhaps some details are different. The Syncnet weights that I had worked for the LRW dataset, but joonsoon's code is faster and more accurate, it would be better to use that.