Question about add_densification_stats

[programath]

I am quite confused that why does the code here use the last value of the last dimension of viewspace_point_tensor.grad to accumulate the absolute value of gradient, while the original gaussian splatting uses the first two value instead? I can't find the explanation from the equation (14) and (15) in the paper.

gaussian-opacity-fields/scene/gaussian_model.py

Line 649 in 98d0858

self.xyz_gradient_accum_abs[update_filter] += torch.norm(viewspace_point_tensor.grad[update_filter,2:], dim=-1, keepdim=True)

[niujinshuchong]

Hi, the third dimension is the sum of the norm of individual pixel gradients (eq. 15 in the paper). You can check the Cuda code here https://github.com/autonomousvision/gaussian-opacity-fields/blob/main/submodules/diff-gaussian-rasterization/cuda_rasterizer/backward.cu#L931-L933 .