search

cleardusk / 3DDFA_V2

The official PyTorch implementation of Towards Fast, Accurate and Stable 3D Dense Face Alignment, ECCV 2020.
MIT License
2.88k stars 510 forks source link

Ground Truth 2D landmarks in 300W-LP dataset #118

Open chuanli11 opened 2 years ago

chuanli11 commented 2 years ago

Hi, I downloaded the 300W-LP dataset and noticed it has a landmarks folder 

ubuntu@ubuntu-desktop:/media/ubuntu/WDC/data/300W-LP/300W_LP$ ls
AFW  AFW_Flip  Code  HELEN  HELEN_Flip  IBUG  IBUG_Flip  landmarks  LFPW  LFPW_Flip

The mat files in that folder contain pts_2d and pts_3d for the augmented faces. Interesting the 2D landmarks are not simply orthogonal projections of the 3D landmarks, but "snapped" to the facial contour as shown in the picture below.

Red: pts_3d Blue: pts_2d

It would be great if you can explain how those pts_2d landmarks are generated. Thank you.

Screenshot from 2022-01-19 10-05-40

cleardusk commented 2 years ago

You may contact Xiangyu Zhu to consult the generation details. I guess the landmarks are generated by warping the 2d landmarks of the original image. If the wrap relation between the orignal image and synthesized image is known, the 2D landmarks can be calculated with the affine operation like grid_sample.

chuanli11 commented 2 years ago

You may contact Xiangyu Zhu to consult the generation details. I guess the landmarks are generated by warping the 2d landmarks of the original image. If the wrap relation between the orignal image and synthesized image is known, the 2D landmarks can be calculated with the affine operation like grid_sample.

Thank you, I will ask Xiangyu.