metric code about cross-identity reenactment

[Harold0530-zhang]

metric code about cross-identity reenactment

[theEricMa]

Hi, thank you for your interest in our code.

• The CSIM is exactly the same as the identity loss in our code which calculates the cosine similarity on the identity embedding extracted from the aligned gt and pred portraits.

• The AKD is the L1 similarity between the 68 facial keypoints extracted from the gt and pred portraits using face_alignment tools.

• The APD and AED are the L1 similarity between the pose and expression coefficients extracted from gt and pred portraints using DeepFaceRecon. We use the original weight, instead of their update in 2023.4. Specifically, the expression coefficients are 64 digits from here, and the pose coefficients are the concatenation of rotation and translation.

• Our FID code adopted the one from pytorch-fid. The metric is computed on the images rendered from camera points in the testset with their ground truths.

For any further questions, please feel free to contact me.