search

hwanhuh / 2D-GS-Viser-Viewer

Simple Viser Viewer for 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
94 stars 6 forks source link

2D disk looks different from the image on the paper. #15

Closed cdfan0627 closed 1 month ago

cdfan0627 commented 1 month ago

Thanks for sharing the viewer. I’m wondering why the 2D disk looks different from the image in the paper.

hwanhuh commented 1 month ago

You can modify the kernel function used to visualize a 2D Disk. Currently, the kernel is defined as follows.

torch.exp(-1/2 * 100 * torch.clamp(opacity-0.5, min=0) ** 2)

I don't know what specific kernel design is used for the paper's disk visualization.

Li-colonel commented 1 month ago

See https://github.com/hbb1/2d-gaussian-splatting/issues/115

hbb1 commented 1 month ago

We should modify the kernel function of 2D Gaussian. https://github.com/hbb1/diff-surfel-rasterization/blob/28c928a36ea19407cd9754d068bd9a9535216979/cuda_rasterizer/forward.cu#L365-L366

float rho = min(rho3d, rho2d);
rho = 100 * max(rho-0.5, 0)

Unfortunately, this requires modifying the Cuda code so that a surfel within a radius is fully opaque.