microsoft / Deep3DFaceReconstruction

Accurate 3D Face Reconstruction with Weakly-Supervised Learning: From Single Image to Image Set (CVPRW 2019)
MIT License
2.18k stars 443 forks source link

About the role of gamma loss #155

Closed HhansssW closed 3 years ago

HhansssW commented 3 years ago

Hi ! Thanks for sharing your great work. I could not understand the role of the gamma loss when using 3 channels illumination (polychromatic light)since you said from your work that the gamma loss is to ensure a nearly monochromatic light. I would really appreciate it if you could give me some information about it. Thanks.

YuDeng commented 3 years ago

Hi, in the original paper we directly set all 3 channels of light to be identical. Later we found that this setting would lead to some artifacts in face textures given input images with abnormal lightings. Therefore, we replace the monochromatic light setting with polychromatic light to alleviate this problem. But we still want the light to be nearly monochromatic to avoid its compensation for face texture. So we add the gamma loss to constrain it.

HhansssW commented 3 years ago

Hi, in the original paper we directly set all 3 channels of light to be identical. Later we found that this setting would lead to some artifacts in face textures given input images with abnormal lightings. Therefore, we replace the monochromatic light setting with polychromatic light to alleviate this problem. But we still want the light to be nearly monochromatic to avoid its compensation for face texture. So we add the gamma loss to constrain it.

Thank you so much : )