search

reshow / PRNet-PyTorch

The training and evaluation code for PRNet (《Joint 3D Face Reconstruction and Dense Alignment with Position Map Regression Network》)
73 stars 16 forks source link

wondering how to accelerate the process of UVMAP2MESH #15

Closed super3kl closed 3 years ago

super3kl commented 3 years ago

cause i want to put the rendering part into the training process,but the rendering process ,especially 'uv2mesh' process takes lots of time. I'll appreciate it if you give me some advice.

reshow commented 3 years ago

The mesh triangles are calculated repeatedly in the functiona, but it is the same for all the meshes. You can pre-compute it and save the result. I think it would be much faster.

super3kl commented 3 years ago

hello and i got some other new questions: 1: have you ever reproducted the papaer's results like mean NME 3.62 ,but i trained your baseline a lot and got around 3.8 NME according the yaw angle. 2: why does gt_pos in dataerloader have to divide by 280, don't know why it goes with 280 as i compute the gt_pos ,and the range is about[-58,312] for 300W-LP,so i am confused about the normalization 3.i would like add your wechat for details communicatio，or just emails chat is fine .here is my e-mail qianqiubiu@163.com

reshow commented 3 years ago
  1. I have not calculate the mean NME of prnet.
  2. Sorry I forget why I choosed 280😂.
  3. rzy503099765@qq.com.
rakacendekia commented 3 years ago

hello and i got some other new questions: 1: have you ever reproducted the papaer's results like mean NME 3.62 ,but i trained your baseline a lot and got around 3.8 NME according the yaw angle. 2: why does gt_pos in dataerloader have to divide by 280, don't know why it goes with 280 as i compute the gt_pos ,and the range is about[-58,312] for 300W-LP,so i am confused about the normalization 3.i would like add your wechat for details communicatio，or just emails chat is fine .here is my e-mail qianqiubiu@163.com

Hello, i got some questions. How is the method to get the mean NME, whether by grouping the database based on its angle, or how?