LansburyCH / relightable-nr

A Neural Rendering Framework for Free-Viewpoint Relighting (CVPR 2020)
Other
116 stars 25 forks source link

Should Relighting gt Loss be added in to the loss function? #9

Open YaelCassini opened 1 year ago

YaelCassini commented 1 year ago

Sorry to bother you, I changed my own data for the test of the relightable-nr model. My data is face mesh, and I have replaced the tex with the diffuse map of the face, but in the training results, the effect of relighting is not very good, the face looks like a metal-like material. After checking the source code, I found The loss of relighting results and relighting ground truth is not added to the loss function, so I wonder if this is the cause of this problem?

LansburyCH commented 1 year ago

Hi YaelCassini,

The original problem setting of our paper takes only images under a same lighting as input, so there is no relighting groundtruth. However, if there is relighting supervision, then you could consider adding relighting losses. For the metal-like appearance, I'm not very certain about this, but you can try increasing the number of sampled directions for the diffuse part and/or removing the specular part.

YaelCassini commented 1 year ago

Thank you very much, I'll try your suggestion then.