3D Gaussian Splatting Renderer implemented in WebGPU (WGPU) and Rust

This code implements a renderer for the reconstructions obtained with 3D Gaussian Splatting

Build

Clone the repository and run

cargo build --release --bin viewer

Run

Use the point_cloud.ply and cameras.json files generated by 3D Gaussian Splatting:

cargo run --release --bin viewer point_cloud.ply cameras.json

Usage

3D Gaussian Splatting Viewer

Usage: viewer [OPTIONS] <INPUT> [SCENE]

Arguments:
<INPUT> Input ply file
[SCENE] Scene json file

Options:
--max-sh-deg <MAX_SH_DEG> maximum allowed Spherical Harmonics (SH) degree [default: 3]
--sh-dtype <SH_DTYPE> datatype used for SH coefficients [default: byte] [possible values: float, half, byte]
--no-vsync
-h, --help Print help
-V, --version Print version

Controls

Use the mouse and WASD + Shift + Space to navigate the camera.

If a scene (cameras.json) file is provided:

• The keys 0-9 can be used to select the views 0-9
• Pressing the R key will select a random view
• Next/Previous view can be selected with Page Up/Down
• N snaps to nearest view
• T starts the tracking shot that visits the test views

About

Splat Sorting : We ported the Fuchsia RadixSort to WGPU for sorting the splats on the GPU.

Performance: The renderer reaches >200 FPS on a NVIDIA 3090 RTX and ~130 FPS on a AMD Radeon R9 380 Series (8 years old). Measurements where taken for the bonsai scene at 1200x799 resolution.

TODO

• better camera controls
• make it work in the browser