search

microsoft / Deep3DFaceReconstruction

Accurate 3D Face Reconstruction with Weakly-Supervised Learning: From Single Image to Image Set (CVPRW 2019)
MIT License
2.16k stars 441 forks source link

About coefficients of R-Net #183

Open hhuyujie opened 2 years ago

hhuyujie commented 2 years ago

Thanks for your released code! In your paper, there are 239 coefficients, but in your code, there seems to be 260. what do the 21 extra coefficients represent? And is there any pattern in the range of these coefficients?

YuDeng commented 2 years ago

In the original paper we assume that the lighting is white light so it only has 9 dimensional coefficients. In this repo we use color light instead so it contains 9*3(RGB) = 27 coefficients.

The coefficients for shape, texture, and expression should have a scale around 1. The rotation is in rad, and the translation is around a scale of 1 in decimeter. The lighting coefficients do not have a certain range.

hhuyujie commented 2 years ago

However, 239 + 9*2 ≠ 260, there are 3 coefficients unmentioned.