Questions about training configuration

[yjhong89]

Dear authors, I am YJHong and thanks for great work!

I am currently training DECA based on this repository and paper.

I preprocessed all of datasets (vggface, bupt, voxceleb2) as mentioned in paper and did landmark cleaning. But my trained model performance is always little bit lower than provided official model (In NOW evaluation)

Could you give me any advice about DECA training ? (Actually I wonder there is another configuration set for traininig official model)

Thank you

Best regards, YJHong.

[oxygenman]

Dear YJHong, could you share your dataset preprocessing and cleaning code?

[yjhong89]

I use voxceleb2, bupt, vggface2

• Get FAN landmarks and skinmask
• Did landmark preprocess (cleaning) for every image as paper mentioned.

[oxygenman]

OK, thank you !

[sangeetajha]

I am trying to train the model with VGGFace2 dataset. I have few questions. YJHong89 can you please help to answer

1. For skinmask - I am using https://github.com/YuvalNirkin/face_segmentation and I am stuck with the installation process itself. Is there a Python package that I can install and generate skin mask ?

2. How did u generate vggface2_train_list_max_normal_100_ring_5_1_serial.npy file ? Can u please share the script ? Since VGGFace2 has so many images per person, I am not sure how to select only five.

Thanks.

[yjhong89]

@sangeetajha

1. I use another skin segmentation mask. Acturally if rendering works well, uv_mask would work without skinmask
2. I've used all images per person, not actually select 5 images per person randomly.

[Brown2Wang]

@yjhong89 Could you please share your data or code for training? Thank you very much

[prokaryote-hgy]

@yjhong89 Hi, how do you generate skin segmentation mask? Could you share the repo you used? Thank you!