Question about the APD (average pose distance) metric calculation

[fusheng-ji]

Hello,

I have a question about APD metric calculation in the paper. As I read the issue in OTAvatar, they compute the APD with the L1 loss on concatenation of roation and translation.

Here is my implement:

First I extract angles and translation coeffs from Deep3DFaceRecon:
angles = pred_coeff[224:227] # euler angles for pose
translation = pred_coeff[254:257] # translation
pose_coeff = np.concatenate([angles, translation], 0)

Then I calculate L1 loss between pose_coeff of predction and ground truth.
apd_loss = np.mean(np.abs(pred_pose_coeff - gt_pose_coeff))

However, after I getting the average metrics on HDTF and VFHQ evaluation, the values of APD are much lower than number in paper.

Could you help me with that?

Many Thanks

[weihaosky]

If I remember correctly, APD metric is only for angles.

[fusheng-ji]

Hello @weihaosky,

Thanks for your clarification. I will compute it with only angles.

Really appreciate your help.