Open slulura opened 3 years ago
hi there, I think it might be helpful to use patching if u wanna calculate the loss in terms of the semantical meaning (i.e., style loss) instead of MSE. FYI, I am implementing a patch-based optimization scheme for NeRF, and the loss is calculated per patches instead of random rays across all images.
@songrise hey, I also implement a patch-sampling-based optimization scheme. For example, N_rand=1024, so I sample a 32x32 patch in each training iteration. But I found the rgbloss was so unstable. It seems that the url you provied has been invalid, I wonder what's your implementation and whether it works.
Would it benefit to use some form of patch sampling in order to include convolutional blocks?