vis results

[Jasper-sudo-Sun]

Thank you for open-sourcing the code. I'd like to ask why, in your visualization results, most of the gaussians still maintain a spherical shape and haven't changed significantly?

[Huijie-Liu]

I have a similar question regarding the optimization of Gaussians. After saving and visualizing the optimized Gaussians as PLY files, I observed that all the Gaussian spheres are almost identical in size. If the primary purpose of these Gaussian spheres is to represent occupancy, wouldn’t using just spherical shapes suffice? This could potentially reduce the dimensionality of parameters.

Below is the visualization of the Gaussians generated based on the 144,000 model (with all Gaussians of category 17, which are empty Gaussians, removed) and corresponding occupancy.

[huang-yh]

Indeed, the Gaussians in the current version do not show much diversity in their scales. We think it is because we have constrained the Gaussians to have similar size as a single voxel, in which case the occupancy supervision itself does not show much diversity in shapes. However, we do notice some flattened Gaussians on the surface of the road.

[bqm1111]

I have a similar question regarding the optimization of Gaussians. After saving and visualizing the optimized Gaussians as PLY files, I observed that all the Gaussian spheres are almost identical in size. If the primary purpose of these Gaussian spheres is to represent occupancy, wouldn’t using just spherical shapes suffice? This could potentially reduce the dimensionality of parameters.

Below is the visualization of the Gaussians generated based on the 144,000 model (with all Gaussians of category 17, which are empty Gaussians, removed) and corresponding occupancy.

How do you generate visualization like this?

[bqm1111]

Indeed, the Gaussians in the current version do not show much diversity in their scales. We think it is because we have constrained the Gaussians to have similar size as a single voxel, in which case the occupancy supervision itself does not show much diversity in shapes. However, we do notice some flattened Gaussians on the surface of the road.

Could you elaborate more on how you have constrained the Gaussians to have similar size as a single voxel?