I am trying to render high dimensional features using latent-gaussian-rasterization. In my use case, I don't need rendered colors so I just leave them untouched. I notice that if I only compute loss on rendered features and do backpropagation, the gradient on gaussians' covariance and opacity will be zero.
Is this behavior correct? It seems better to also propagate gradient from feature to covariance & opacity.
Dear authors,
First thanks for this great work!
I am trying to render high dimensional features using
latent-gaussian-rasterization. In my use case, I don't need rendered colors so I just leave them untouched. I notice that if I only compute loss on rendered features and do backpropagation, the gradient on gaussians' covariance and opacity will be zero.Is this behavior correct? It seems better to also propagate gradient from feature to covariance & opacity.
Many thanks!