search

cbsropenproject / 6dof_face

Codes and data for the TIP 2023 paper: Towards 3D Face Reconstruction in Perspective Projection: Estimating 6DoF Face Pose from Monocular Image
50 stars 6 forks source link

Evaluation on BIWI #9

Closed asw91666 closed 11 months ago

asw91666 commented 1 year ago

Hi. I evaluated your model on the BIWI using the code you provided.

Used GT mesh (from obj file) Used pred bbox (from FAN that you referenced in your paper) Used your code for the rotation, translation, and ADD evaluation. The results are as follows:

"mae_r": "4.35 degree", "pitch_mae": "5.20 degree", "roll_mae": "3.59 degree", "yaw_mae": "4.27 degree"

"mae_t": "26.23 mm", "tx_mae": "14.56 mm", "ty_mae": "18.18 mm", "tz_mae": "45.94 mm",

"ADD": "57.50 mm",

The rotation error and translation error are slightly worse than the performance reported in the paper, but I believe this could be attributed to the details during the implementation. However, the ADD I measured is significantly better than what was reported in the paper. Do you have any idea why this might be the case?

cbsropenproject commented 1 year ago

For ADD result, we found a bug in our code, we used a random model which should be the GT model. We forgot to change it. For other metrics, the results should be similar to ours. Please check the RGB images and transformed GT pose according to ARKitFace dataset.

asw91666 commented 1 year ago

Ok Thanks !

cbsropenproject commented 1 year ago

We have updated our trained model. Please download the newest model from the link in readme.