Dear ShenhanQian,
i have a question regarding cross reenactment. I observed quite noticable identity spill on my data. Animating the face of person 1 with person 2's motion sequence does introduce a bit of person 2's face geometry. Is this something you have experience with? Is there a way to disentangle face geometry and expression when cross reenacting? I read the flame paper and they state that expression transfer should be identity independent, but that is not the case in my experience. You can't have identity independence when tracking a subject, since tracking a person with neutral expression does NOT produce a neutral FLAME model, any facial geometry is already reflected in the combination of expression and shape parameters. If i track my neutral face and then set expression to zero, my avatar does not have my neutral expression. And if i put the "neutral" face of another person on my face, it will not be neutral, because of their face geometry.
Do you know how to reduce this identity spill or have any ressources on research about disentangling expression from identity? What where your experiences during your experiments?
Dear ShenhanQian,
i have a question regarding cross reenactment. I observed quite noticable identity spill on my data. Animating the face of person 1 with person 2's motion sequence does introduce a bit of person 2's face geometry. Is this something you have experience with? Is there a way to disentangle face geometry and expression when cross reenacting? I read the flame paper and they state that expression transfer should be identity independent, but that is not the case in my experience. You can't have identity independence when tracking a subject, since tracking a person with neutral expression does NOT produce a neutral FLAME model, any facial geometry is already reflected in the combination of expression and shape parameters. If i track my neutral face and then set expression to zero, my avatar does not have my neutral expression. And if i put the "neutral" face of another person on my face, it will not be neutral, because of their face geometry.
Do you know how to reduce this identity spill or have any ressources on research about disentangling expression from identity? What where your experiences during your experiments?