search

anhttran / 3dmm_cnn

Regressing Robust and Discriminative 3D Morphable Models with a very Deep Neural Network
Other
857 stars 254 forks source link

How to train the model and define the label？ #13

Closed ChuckGithub closed 7 years ago

ChuckGithub commented 7 years ago

It's very kind of you to share the model. I'am a new guy to 3D face reconstruction and I want to train my own model, what should i do? Looking forward to your reply.

anhttran commented 7 years ago

Sorry for my late reply. You can find our description of how to get training labels in section 3.1, and how to train the network in section 3.2: https://arxiv.org/abs/1612.04904

Best, Anh

ChuckGithub commented 7 years ago

Thank you for your answer, but I have another question is how to use the generated 3DMM model as a label and on behalf of this category. That is to say which 3DMM parameters are used as a training label.

anhttran commented 7 years ago

We used 99 shape parameters + 99 texture parameters. They form 198D vectors, which were used as training labels.

ChuckGithub commented 7 years ago

OK,Thank you,This information can be obtained from your paper, how to choose the 198D vectors from the 3DMM model,and what are the specific requirements for these parameters.

anhttran commented 7 years ago

For each CASIA image, we estimate 3DMM parameters, using the traditional method. Each parameter is a floating number in range [-3, 3]. Then, we computed weighted average of these parameters from all images of the same subject, and use the pooled parameters to form a 198D ground-truth label.

In CNN training, we used 1 LMDB to store the images, and 1 LMDB to store the corresponding labels.

ChuckGithub commented 7 years ago

OK,Thank you,This process is clear.

WWWJL commented 5 years ago

OK,Thank you,This process is clear.

hi,can you train about your own dataset ?