High Performance Point Cloud Rendering

[benjaminrwilson]

Hi,

What is the most performant way to render large point clouds (100,000+ points)? I've had success by storing the points in Mesh::new(PrimitiveTopology::PointList), but I haven't found a way to change the point size or to have per point coloring.

Is this possible with this approach? If not, what would it require?

Thanks!

[torsteingrindvik]

I'm not sure about most performant.
But I think you'll be able to achieve this by using a storage buffer written on the CPU and copied and bound in the GPU.
If there is 1 color per point and 1 vertex per point you should be able to use the vertex index directly to read the correct color in the shader.

[superdump]

I did some hacks here for drawing many quads: https://github.com/superdump/bevy-vertex-pulling/blob/main/examples/quads/main.rs The approach is basically what @torsteingrindvik was suggesting.

[benjaminrwilson]

Awesome. Thanks @torsteingrindvik and @superdump! The advice, example, and https://github.com/bevyengine/bevy/blob/main/examples/shader/shader_instancing.rs got me what I needed.

[ottah]

If you want really performant point clouds you could write a compute rasterizer and output to a texture.

Theres some linked papers in this repo https://github.com/m-schuetz/compute_rasterizer

[benjaminrwilson]

Thanks, I'll have a look!

[geoproces]

Hi Benjamin , I am land surveyor working with points clouds ,is there any chance you can send me example of point clouds to be imported tu bevy ?

[mosure]

see https://github.com/mosure/bevy_gaussian_splatting

[geoproces]

Thx. mousure , however I am using Lidar scarer ,not a camera , the data are process and calculate in hardware soft . I just need display them in Bevy . The gaussian_splatting is very promising , but as i sett i do not have any pictures . Any ideas ?. Best Regards

[hacknus]

In case this is useful for anybody: I created a crate for a voxel plot, mainly built around the custom shading example from bevy, but with transparency support: https://github.com/hacknus/bevy_voxel_plot

Can easily handle 1000000 voxels.

[rlamarche]

Hi, I developed a point cloud plugin for Bevy : https://github.com/rlamarche/bevy_pointcloud
It's the beginning and I'm using shader instancing instead of a storage buffer.

I'm not experienced enough to tell what is the most efficient way to do it, but I'm quite satisfied by the fact that my implementations uses Bevy's frustum by computing AABB.

[hacknus]

Cool! Does it support alpha lower than 1?

[rlamarche]

Not yet. First I would like to implement high quality rendering and eye dome lightning from this paper : https://www.cg.tuwien.ac.at/research/publications/2016/SCHUETZ-2016-POT/SCHUETZ-2016-POT-thesis.pdf

For transparency, it depends what is the objective. It could be an addition of all splats, then normalization.
Or "depth peeling" that would need N passes (slow) depending on the complexity of the point cloud.

For example, if the point cloud is from LIDAR or photogrametry, we only have the surfaces. But if you're looking for something to see "inside" the point cloud, it's more complicated, and normalization would be a good alternative.

Anyway, I will give a try for all these techniques, because I'm learning shaders, and it is a very good exercise !

[hacknus]

Ah, cool! Yeah, there exist some other projects doing AABB with bevy and point clouds/voxel plots ( e.g. https://github.com/ForesightMiningSoftwareCorporation/bevy_aabb_instancing, however this is not up to date with bevy afaik). But I think AABB does not support alpha < 1 in general. My bevy_voxel_plot project supports alpha < 1 but it is less efficient and struggles with more than 1 million datapoints. So if you manage to make some progress in that direction let me know!

My objective is to render 3D scans, where I map the intensity of a pulse (reflected) to the opacity. This let's me visualize layers/reflections inside the object.