Configuration of FG2018 evaluation

[yjhong89]

Dear authors, Hi. I am YJ Hong and thanks for sharing great work!

I checked constructing 3d meshes for NOW validation dataset part in the official code but there is no related code for FG2018 evaluation.

I implemented my own FG2018 evaluation script and run with provided trained model, and evaluation result is slightly better than that of paper.

In an order of (median, mean, std) [Paper] HQ (1.45, 1.89, 1.68), LQ (1.48, 1.91, 1.66) [Provided checkpoint] HQ (1.32, 1.71, 1.5), LQ (1.46, 1.91, 1.7)

I neutralized pose and expression as NOW validation and crop images to have 224 sizes. The only difference is the scale arguments since I used default 1.25 scale argument for "TestData" class.

May I ask you to provide the exact configuration of constructing 3D mesh of FG2018 evaluation, so can get the same result with paper ?

Thank you. YJHong.

[ccyyatnet]

hi @yjhong89 , Could you share your choices of configuration? Thks

[yjhong89]

@ccyyatnet My configurations are

1. Make object file of fg2018 images (Subset_2D_FG2018 2000 images (HQ 656, LQ 1344)) using TestData class, same crop and scale arguments.
2. Compute distances referenced by fg2018 official repository issue.
3. Get same 7 landmarks to rigidly align and neutralize expression and pose vector (set to 0) similar with this code

Those are all my configurations.