RVizSplat

RVizSplat is an RViz2 display plugin that provides end-to-end visualization of 3D Gaussian Splats in RViz.

Build status

       

How to build and run from source

mkdir -p ~/ros_ws/src
cd ~/ros_ws/src
git clone https://github.com/RVizSplat/RVizSplat.git
cd ~/ros_ws
rosdep update
rosdep install --from-paths src --ignore-src -r -y
colcon build --symlink-install --cmake-args -DCMAKE_BUILD_TYPE=Release
source install/setup.bash

Installation via apt

This feature is currently under development (See ros/rosdistro#50909 for details)

After sourcing your ROS 2 environment:

sudo apt-get install ros-$ROS_DISTRO-gsplat-rviz-plugin ros-$ROS_DISTRO-gsplat-publisher ros-$ROS_DISTRO-gsplat-msgs

Examples

Rendering of a LeRobot SO-100 arm with a scene containing 6 million splats

Rendering of transparent markers along with other gaussians

Using OIT for performance optimization

If you have a resource constrained CPU and a weaker GPU (just integrated graphics), you might want to consider bypassing sorting entirely. For this use case, we provide OIT based implementations.

To activate this, follow the "Advanced" options in the RViz plugin and select WBOIT.

Architecture

Coming soon!

Evaluation

The gsplat_plugin_evaluation/eval.py script computes image quality metrics (PSNR, SSIM, LPIPS) between a ref_folder and an eval_folder.

Images are matched by the trailing 3-digit number in the filename (e.g. img_001.png in the ref_folder is paired with *_001.png in the eval_folder).

Usage

cd gsplat_plugin_evaluation
python eval.py <ref_folder> <eval_folder> [--metrics psnr ssim lpips] [--lpips-net alex|vgg]

Argument	Description
ref_folder	Folder containing reference (ground-truth) images
eval_folder	Folder containing images to evaluate
--metrics	Space-separated list of metrics to compute (default: all three)
--lpips-net	Backbone network for LPIPS — alex (default) or vgg

Examples

Compute all metrics using the default AlexNet backbone:

python eval.py data/ref data/eval

Compute PSNR and LPIPS with VGG backbone:

python eval.py data/ref data/eval --metrics psnr lpips --lpips-net vgg

Output

The script prints a per-image table and a mean row at the bottom:

Image                    PSNR        SSIM       LPIPS
--------------------------------------------------------
img_001.png           32.1500      0.9210      0.0431
img_002.png           29.8300      0.8970      0.0612
--------------------------------------------------------
Mean                  30.9900      0.9090      0.0522