search

city-super / Scaffold-GS

[CVPR 2024 Highlight] Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering
https://city-super.github.io/scaffold-gs
Other
638 stars 47 forks source link

Question about grid_scaling #58

Open MuQoe opened 1 month ago

MuQoe commented 1 month ago

upon this code section

` # post-process cov scaling = scaling_repeat[:,3:] torch.sigmoid(scale_rot[:,:3]) # (1+torch.sigmoid(repeat_dist)) rot = pc.rotation_activation(scale_rot[:,3:7])

# post-process offsets to get centers for gaussians
offsets = offsets * scaling_repeat[:,:3]
xyz = repeat_anchor + offsets

`

What does this stands for scaling_repeat[:,3:], i have found that this is from grid_scaling initialize by

scales = torch.log(torch.sqrt(dist2))[...,None].repeat(1, 6) Could you explain the role of each dimension in scaling_repeat, especially in the context of the slicing [:,3:] and [:,:3]?

Thanks

inspirelt commented 4 weeks ago

The [:,:3] controls the step size of offset. The [:,3:] serves as the base scale for neural gaussian's shape, which means the cov MLP learn a residual scales.