splatviz

See some example editings in this video.

NEW UPDATE: splatviz now supports live tracking of a running 3DGS training: video.

Install

git clone https://github.com/Florian-Barthel/splatviz.git --recursive

conda env create \
-f environment.yml

conda activate gs-view

micromamba env create \
--file environment.yml \
--channel-priority flexible -y

micromamba activate gs-view

Launch

python run_main.py

or launch with a specified path to look for .ply or .yml files:

python run_main.py --data_path=path/with/ply/files

python run_main.py --mode=attach

python run_main.py \
--mode=decoder \
--data_path=path/with/pkl/files \
--ggd_path=path/to/gaussian_gan_decoder/project

In some cases you will have to add this variable so that opengl uses the correct version: export MESA_GL_VERSION_OVERRIDE=3.3

Widgets

Once you run the run_main.py file, the viewer will directly load the first scene that is found in the data directory you have specified. You can change the scene by clicking the Browse button. You will be displayed a list of available .ply files (or .yml files that correspond to compressed gaussian scenes using this compression method). If you have a lot of scenes to chose from, you can also use the Filter textfield providing comma separated keywords (eg. lego,30000,baseline). You will then only see those scenes that contain all keywords.

Additionally, you can also view multiple 3D scenes at once. Simply click Add Scene, which loads the same scene as before and then change the respective scene to another .ply file using the new browse button. The scenes are either rendered next to each other or in a split screen mode when activating the Splitscreen checkbox.

You can download some example scenes here. They were created using the gaussian splatting compression method from: https://fraunhoferhhi.github.io/Self-Organizing-Gaussians.

The edit widget is the core functionality of this 3D viewer. It allows for real time editing of the gaussian python object during runtime. The code that you type in this text area will be executed just before the gaussian object is forwarded to the cuda renderer. This means, the editing capabilities are unlimited. As long as the provided code is executable, you can type any kind of python code and also import new libraries. An example could look like this, where all scales are set to -8 (before activating with exp()) and all opacity values are set to 10 (before activating with sigmoid()), while also the background is set to 1, which corresponds to white.

gaussian._scaling = gaussian._scaling * 0 - 8
gaussian._opacity = gaussian._opacity * 0 + 10
self.bg_color[:] = 1

To enable smooth editing transitions, you can create sliders (press Add Slider) which you can access in the editor text by typing slider.name (eg. slider.x). An example could look as follows. Here, we create a boolean mask that filters all gaussians that are smaller than the value stored in slider.x.

mask = torch.linalg.norm(gaussian._scaling, dim=-1) < slider.x

gaussian._xyz = gaussian._xyz[mask]
gaussian._rotation = gaussian._rotation[mask]
gaussian._scaling = gaussian._scaling[mask]
gaussian._opacity = gaussian._opacity[mask]
gaussian._features_dc = gaussian._features_dc[mask]
gaussian._features_rest = gaussian._features_rest[mask]

Lastly, you can save and load presets of code snippets so that you don't have to type the same code again after closing the application. Those presets are stored in a .json file (presets.json).

The evaluate widget can be used to debug the gaussian splatting object. By typing python code, which is executed after the rendering, you can access any variable from the rendering context and visualize them in a histogram. Some useful variables might be:

• gaussian
• render
• render_cam
• self

You can also access variables that you have defined in the editor of the Edit Widget.

In the camera widget you can define the type and parameters of the camera. Most importantly, you can choose between the two modes Orbit and WASD.

In Orbit mode, the camera is looking at a specific point in 3D space, and you control the pitch and yaw of the camera rotating on a sphere around that point by dragging with the mouse over the rendered view.

In WASD mode, you fly through the scene using the mouse and the WASD keys similar to the camera controls in Unity.

Important: If the loaded scene is rotated incorrectly, you can adjust that by steering the camera so that it is looking straight up or down. Then set the up-vector to the current viewing direction by pressing **Set Current Direction **. If you were looking down, you will have to press the Flip button, since the scene is now upside down.

The video widget creates a video sequence of a full rotation around the current object. Simply define the height of the camera and the rendering resolution. While the video is rendering, the UI is frozen. A loading screen is shown in the terminal output.

🚧This feature is still under construction 🚧. Currently, it is not very intuitive to generate videos, as the camera position is only defined by the Camera Height parameter. This will be changed in the future.

In the render tab, you can specify the rendering resolution (at the moment only squared resolutions) and the rendering mode. Next to RGB, you can also render the image 3D depth and the alpha values.

To save the .ply file or a rendering of the current (edited) 3D scene press the respective button in the save widget. Those files will be saved under _screenshots or _ply_files. The viewer might be frozen for a short while, after saving a ply file.

⭐ Recent Features

Version 1.1.0

• 2024-08-12: Added a new Training Widget to inspect live training stats and to pause training
• 2024-08-11: Attach to a running 3DGS training
• 2024-08-10: Refactor rendering class for easier addition of new renderer
• 2024-08-07: Better Mouse Control (translate with middle mouse button)
• 2024-08-05: Allow editing of existing sliders
• 2024-07-30: Store slider values in presets
• 2024-07-28: New Performance Widget
• 2024-07-28: Editor now highlights special variables (gs, self, slider) and gives tooltips

Version 1.0.0

• 2024-07-12: Rebuild the whole application with imgui_bundle
• 2024-07-05: Compare two or more Gaussian Splatting scenes side by side

🔭 Future Plans

• Test other 3DGS render for OSX support
• Improve the Video Widget
• Tutorial for coding new widgets or renderer

Contribute

You are more than welcome to add further functionality or a better design to this interactive viewer! The main goal is to create an easy-to-use tool that can be applied for debugging and for understanding 3D Gaussian Splatting objects. For reformating code please use black with --line-length 120.

Citation

If you find this viewer useful, please consider citing our work:

@misc{barthel2024gaussian,
    title={Gaussian Splatting Decoder for 3D-aware Generative Adversarial Networks}, 
    author={Florian Barthel and Arian Beckmann and Wieland Morgenstern and Anna Hilsmann and Peter Eisert},
    year={2024},
    eprint={2404.10625},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

References

This viewer is inspired by the visualizer from Efficient Geometry-aware 3D Generative Adversarial Networks (EG3D).

• GUI: pyimgui and
  imgui_bundle which are python wrappers for the c++ library
  ImGUI
• Original code base: EG3D
• 3DGS: 3D Gaussian Splatting
• Compressing 3DGS scenes: Compact 3D Scene
  Representation via Self-Organizing Gaussian Grids
• 3DGS Rasterizer with depth and alpha: Diff
  rasterizer with depth and alpha