search

Lab-of-AI-and-Robotics / GS_ICP_SLAM

[ECCV 2024] RGBD GS-ICP SLAM
MIT License
347 stars 36 forks source link

a methodology flaw? and a bug in render_3? #42

Open tarquai opened 3 months ago

tarquai commented 3 months ago

Hi and thanks for sharing your good work, I would like to work with it more BUT:

There might be a bug in render. when using gsplat and supersplat to render, we get elongated gaussians, look at the other issue in this forum for images: https://github.com/Lab-of-AI-and-Robotics/GS_ICP_SLAM/issues/27 After an investigation, the gaussians scale has a very small z component compared to the x,y components. The "render" and "render_3" function process the Gaussians differently (they contain custom code).

Since gaussian splatting and rendering is well-defined. it should work with "regular" off-the-shelf gaussian splatting visualisation libraries.

Thanks.

tarquai commented 2 months ago

look at the link to the other issue. it's just a different convention for the quats (w,x,y,z) / (x,y,z,w)