DECA
DECA copied to clipboard
yfeng95
Prepare dataset
Can you please give me more explanation about preparing data for training the model from scratch? I want to use VGGFace2 for the first step. Do I need to generate landmarks for every image and save the landmarks as a text file or npy file? and what about masks? I really appreciate it if you give me some clue
I am trying to do this step: b. Prepare label FAN to predict 68 2D landmark face_segmentation to get skin mask
How do you get the five faces of every subject? (vggface2_val_list_max_normal_100_ring_5_1_serial.npy') I could get landmarks or skin mask. Author choose five normoal faces with every identity. It is more important!
@littlePrince126 Can you please help me on generating facial masks? I couldn't install https://github.com/YuvalNirkin/face_segmentation
when I run the repository I get this error
@littlePrince126 I want to try with some simple data just to run the code so I didnt get the five faces from every object yet. we can work on it together if you want.
@Ned09 It seems that you don't set the variables in a right way. You can have a try by using the python script.
@littlePrince126 Do you mean without installing this repository and just by running the python script in the "face-segmentation/interfaces/python/face-seg.py" , I can generate the segmentation images?
@Ned09 Yes, because you should compile caffe and opencv if you want to install this repo. Anyway, you need a caffe environment at first.
@littlePrince126 Hi, what's your GPU used when runing face-segmentation? when I use caffe-gpu env with docker,
F1103 05:56:53.555362 430 pooling_layer.cu:212] Check failed: error == cudaSuccess (8 vs. 0) invalid device function
I got this error, It may because the difference of GPU computation ability between authur's and mine.
What's your situation?
identity. It is more important!
Have you got the file(vggface2_val_list_max_normal_100_ring_5_1_serial.npy)? I would really appreciate it , if you share it with me.
@1180800817 I generate it by myself. You can generate it with dataset（VggFace2，BUPT-Face，voxceleb2 ）. voxceleb2 is a video dataset, you should choose good quality videos first, then get the images from videos.
@littlePrince126 Can you reproduce the reported result on NoW using your generated data? Additionally, could you please share vggface2_val_list_max_normal_100_ring_5_1_serial.npy with me? Even a demo is helpful! I just want to know the explicit format of this file.
@lhyfst I trained the model without any pretrained model. Finally, I get median1.217mm mean1.517mm std 1.249mm with NOW validation set. (DECA median:1.18, mean:1.46mm,std:1.25mm)
I've test DECA pretrained model in NoW dataset, result is median:1.15, mean:1.44mm,std:1.24mm... but in DECA paper they report median:1.09, mean:1.38mm,std:1.18mm @littlePrince126 朋友，方便加个微信讨论吗？
@Shelomith Because your results is based on validation set. **(median:1.09, mean:1.38mm,std:1.18mm )**This result is based on test set. But I only get results(meadian:1.1778,Mean:1.4636,std:1.2529) with DECA model.
meadian:1.1778,Mean:1.4636,std:1.2529 This result is the same as papr(DECA)
I've test DECA pretrained model in NoW dataset, result is median:1.15, mean:1.44mm,std:1.24mm... but in DECA paper they report median:1.09, mean:1.38mm,std:1.18mm @littlePrince126 朋友，方便加个微信讨论吗？ hello 你好 最近也在跑这个代码 方便加个微信讨论吗？或者邮箱？
@littlePrince126 Hi,can you share the method of choosing five normoal faces?Thank you!
@lhyfst I trained the model without any pretrained model. Finally, I get median1.217mm mean1.517mm std 1.249mm with NOW validation set. (DECA median:1.18, mean:1.46mm,std:1.25mm)
您好，方便给个联系方式吗？想请教一下怎么训练的，有偿。
choose good quality videos
Hi, how do you choose good quality videos? @littlePrince126
你能帮我制作面膜吗？我无法安装 https://github.com/YuvalNirkin/face_segmentation
当我运行存储库时，出现此错误
May I ask how you resolved this issue and ultimately achieved training?
How do you get the five faces of every subject? (vggface2_val_list_max_normal_100_ring_5_1_serial.npy') I could get landmarks or skin mask. Author choose five normoal faces with every identity. It is more important!
Hi！Can you reproduce the model training？
