cleardusk/3DDFA

Regressor from 3D dense vertices to 3D keypoints

FishWoWater opened this issue · 8 comments

FishWoWater commented

Hi! Thanks for your nice work!

I am wondering how to obtain the indices of 68 keypoints from original 5w+ vertices

I want to get the indices of 106 keypoints, is there any solution?

Thanks!

cleardusk commented

You may annotate it by yourself.

FishWoWater commented

@cleardusk Is there any tool I can make use of? (e.g. for visualization?)

cleardusk commented

meshlab

FishWoWater commented

@cleardusk
Hi! I have another question:

FishWoWater commented

If I want to re-train 3DDFA with 128x128 inputs. I need to

  1. resize input images to 128 x 128
  2. decompose f from 3x4 camera parameters ( by pow(np.linalg.det(cam[:, :3]), 1/3) )
  3. decompose offset from camera matrix
  4. let f' = f * 128 / 120 and t_x = t_x + 4; t_y = t_y + 4;

right?

FishWoWater commented

If I want to re-train 3DDFA with 128x128 inputs. I need to

  1. resize input images to 128 x 128
  2. decompose f from 3x4 camera parameters ( by pow(np.linalg.det(cam[:, :3]), 1/3) )
  3. decompose offset from camera matrix
  4. let f' = f * 128 / 120 and t_x = t_x + 4; t_y = t_y + 4;

right?

I visualized the images in 120x120 and their 128x128 counterparts and I believe it's ok

  • 👍3