search

hbb1 / 2d-gaussian-splatting

[SIGGRAPH'24] 2D Gaussian Splatting for Geometrically Accurate Radiance Fields
https://surfsplatting.github.io
Other
1.86k stars 113 forks source link

How can I recover the covariance matrix using quaternions and two-dimensional scales? #88

Open SlamCabbage opened 2 months ago

hbb1 commented 2 months ago

padding zeros to the scales and pass them into build_scaling_rotation

SlamCabbage commented 2 months ago

torch.exp is used in build_scaling_rotation, so is the "scales" saved at the end the actual size of the Gaussian in this scene?

hbb1 commented 2 months ago

Hi, when I refers to the scales, I mean the post-activation scales. Yes, the scales are indeed measured in the actual scene; that is, the world space. You need to append the scale-post-activation for recovering a full covariance.

SlamCabbage commented 2 months ago

image 大佬您好，我用中文了，这是waymo数据，我可视化它的协方差，这个是按照scale的0.05倍可视化的。如果按照1.0倍进行可视化，那么整个场景都是一坨纠缠在一起的，完全看不出来。

DavidXu-JJ commented 2 months ago

Waymo should not be a problem. https://streetunveiler.github.io/ I use 2DGS in my project above. And I at least successfully reconstruct the scene.

SlamCabbage commented 2 months ago

image 我对scales做了一个exp操作之后，可视化的结果如上图。 用2d gs可视化工具能看到这个场景的地面点虽然不多，但是在图像上是覆盖全了，为什么我对高斯进行可视化，地面并没有能够把空出补全的大高斯平面呢？

SlamCabbage commented 2 months ago

Waymo should not be a problem. https://streetunveiler.github.io/ I use 2DGS in my project above. And I at least successfully reconstruct the scene.

Already following, looking forward to open source