GraphDeco Viewer

GraphDeco Viewer is a modern, Python-based 3D visualization tool designed as a replacement for the legacy SIBRViewer, developed at INRIA Sophia-Antipolis. It provides a flexible and lightweight platform for visualizing outputs of 3D algorithms, making it particularly well-suited for research, development, and prototyping in computer graphics and vision. The viewer is written completely in Python to allow for easy installation and prototyping. It is based upon ImGui and uses imgui-bundle for the Python bindings.

NOTE: The current release is in alpha mode, and there may be breaking changes in future updates as the project evolves.

Installation

The package can be installed using PIP as follows:

pip install git+https://github.com/graphdeco-inria/graphdecoviewer

Citation

If you use this software in your publication, please consider citing it:

@software{shah_graphdecoviewer,
  author = {Shah, Ishaan and Meuleman, Andreas and Lanvin, Alexandre and Drettakis, George},
  title = {GraphDeco Viewer},
  url = {https://github.com/graphdeco-inria/graphdecoviewer},
}

Usage

The viewer can run in three different modes:

• LOCAL: Opens a window on the same device.
• SERVER: Creates a websocket server on the device which would perform all of the computation to be displayed on another device.
• CLIENT: Opens a window to display the GUI and connects to the websocket server.

Creating a new Viewer

To create a new viewer, you must inherit from the Viewer class and override the following function:

• __init__: If you override this function, you must necessarily call the __init__ of the parent class using super().__init__(mode) to ensure the state variables and appropriate modules are imported.
• create_widgets: Overide this function to define widgets required by the viewer along with any additional state variables.
• import_server_modules: Import any modules required only on the server here. This typically would involve packages such as torch and other local code dependancies which might not be available in the client environment. You should assume the modules imported here are only available in server_(send|recv) and step functions.
• step: Define any computation in this function. This function will only be called on the server in network mode
• show_gui: Define the GUI in this function. This function will only be called on the client in network mode.
• (client|server)_send: Override this function to send any data from the client or server. The function should return a tuple with the first element being any binary data, and the second being a dictionary containing any text metadata. Either elements can be None if the viewer doesn't need to send that type of data. The default implementation returns (None,None).
• (client|server)_recv: Override this function to define how to update the viewer state with received data. The functions is expected to receive two arguments, the first being the binary data and, the second being the text.

Creating new Widgets

To create a new widget, you must inherit from the Widget class and override the following functions:

• __init__: If you override this function, you must necessarily call the __init__ of the parent class using super().__init__(mode) to ensure an appropriate ID gets assigned the the widget which is required for correct functioning in network mode.
• setup: This function will be called after initializing glfw and OpenGL. Setup related to the above two should go here (For eg. creating an OpenGL Texture).
• destroy: This function will be called when the application exits. Override this function to deallocate any OpenGL objects or anything else which doesn't support automatic garbage collection.
• step: Same as Viewer.step.
• show_gui: Same as Viewer.show_gui.
• (client|server)_send: Same as Viewer.(client|server)_send
• (client|server)_recv: Same as Viewer.(client|server)_recv

Widgets

Following is the list of widgets which are available/planned to be included in the viewer by default.

Name	Local	Network	Description	Note
NumpyImage	x	x	Displays a np.ndarray as a 2D Texture
TorchImage	x	x	Displays a torch.Tensor as a 2D Texture w/o requiring a roundtrip to the CPU	Requires cuda-python
PerformanceMonitor	x	x	A widget to display timings for different phases of your algorithm
FPS Camera	x	x	A first person camera
EllipsoidViewer	x	x	A widget to visualize gaussians as Ellipsoids
RadioPicker	x	x	A simple radio selection widget
EllipseViewer			A widget to visualize 2D gaussians as Ellipses
PixelInspector			A widget which allows to zoom onto a texture and inspect pixels
ImageCompare			A image slider to compare two images side by side