Question about gradients from semantic supervision

[turandai]

Hi, Thanks for opening source of this wonderful work. I met an issue when reading your code. You mentioned in the paper that the spatial property (geoemtry) of the gaossians are also affected by the semantic supervision. However, this part of code seems only propagating the gradients from semantic supervision to the semantic features itself, the alpha and other gaussian properties remains unchanged. If I understand this right, is there any reason for this choice? Thx!