Problems with the Deca training dataset

[nainsheng]

Hello author, I am currently studying this paper and debugging the code. The demo has been successfully debugged, but the details of the training code are not clear, as shown in the following part:
b. Prepare label
FAN to predict 68 2D landmarks
face_segmentation to get skin mask

I would like to ask how this part should be operated, can you give more details?

[xuduo18311199384]

Train Datas include "image.png + landmarks.npy(FAN detect) + segmentation.npy(face_segmentation detect)"

[nainsheng]

@xuduo18311199384
Thank you for your reply. The data related to FAN in the first part has been processed, but we encountered a problem in the facial segmentation step. Firstly, this link needs to install the caffe framework, but many methods have been used but cannot be successfully installed. Secondly, I noticed that I still need Vggface2_train_list_max_normal_100_ring_5_1serial.npy files. Can you give me some prompts? Thank you very much for your help.

[xuduo18311199384]

1. face_segmentation part. You must install caffe(with gpu).

[nainsheng]

Thank you for your reply. I have completed facial segmentation. What should I do about Vggface2_train_list_max_normal_100_ring_5_1serial.npy? Is it the best 5 facial images for each candidate in the dataset?

[xuduo18311199384]

The follow picture should be made by Vggface2_train_list_max_normal_100_ring_5_1serial.npy
each candidate : image/225/0000383.png, 68kpts/225/0000383.npy, face_segmental/225/0000383.npy,

[shape_consistent_loss & detail_loss] require images >= 3 for each candidate

[anushka17agarwal]

Hey, I had a doubt. In the vox.py file (Line 126- 128)
images_array = torch.from_numpy(np.array(images_list)).type(dtype = torch.float32) #K,224,224,3 kpt_array = torch.from_numpy(np.array(kpt_list)).type(dtype = torch.float32) #K,224,224,3 mask_array = torch.from_numpy(np.array(mask_list)).type(dtype = torch.float32) #K,224,224,3
The shape of kpt_array is 224, 224, 3 , but the face alignment library returns a shape of (62, 3). So, how exactly were these keypoints transformerd

[zhumeichoubao]

Hey, I had a doubt. In the vox.py file (Line 126- 128) images_array = torch.from_numpy(np.array(images_list)).type(dtype = torch.float32) #K,224,224,3 kpt_array = torch.from_numpy(np.array(kpt_list)).type(dtype = torch.float32) #K,224,224,3 mask_array = torch.from_numpy(np.array(mask_list)).type(dtype = torch.float32) #K,224,224,3 The shape of kpt_array is 224, 224, 3 , but the face alignment library returns a shape of (62, 3). So, how exactly were these keypoints transformerd

I have the same question, how do you solve it. Can you give me some prompts? Thank you very much for your help.

[jylovec]

The follow picture should be made by Vggface2_train_list_max_normal_100_ring_5_1serial.npy each candidate : image/225/0000383.png, 68kpts/225/0000383.npy, face_segmental/225/0000383.npy, [shape_consistent_loss & detail_loss] require images >= 3 for each candidate

why [shape_consistent_loss & detail_loss] require images >= 3 for each candidate？I would appreciate it if you reply me。

[emlcpfx]

Were you able to train the model?