Faceswap Forum

For example, if the original video A & B are both a person holding a mic singing. After extracting the pics, there will be some pic that the faces (both A&B) are partially blocked by mic or the other waving hand. Other examples like people are eating, and there will have some sending-food-to-mouth pics. Should I still use these pic for training to show the eating actions? Or I only need those pics which include the full face without blocking from different angles?

Thanks.

From the advice I've read from more experienced people, and the little experience I've acquired myself, the answer is "it depends" (which is the answer you'll get the most).

Ideally, you want clear, quality faces with no obstacles, unless you're going to swap to another face that is being blocked in a similar way. So, if you're planning to transfer person B's face onto someone who is using a microphone, then it's not a bad idea to feed the model with situations similar to the ones it's going to be tackling. So long as those faces are shot from similar angles. This is a general rule, by the way. There's no use in feeding the model tons of top-down faces if you want it to solve only bottom-up faces.

On the other hand, even if your target isn't someone using a microphone or eating hamburgers, you might still want to include those faces if they're shot from angles (or facial expressions) you'll need swapped but you don't have enough of.

Personally, I remove obstructed faces unless I know that they have been properly masked for obstructions.

If they have been properly masked for obstructions, then I would include them.

Ultimately, the question is, what do you want the model to learn? A face or a microphone? It's the face, and feeding it images of microphones won't help it any.