K3D-jupyter is a widget which primarily aims to be an easy and efficient 3D visualization tool. It focuses on straightforward user experience, providing API similar to well known Matplotlib. On the other hand it uses the maximum of accelerated 3D graphics in the browser. Those features make it a lightweight and efficient tool for diversity of applications. Some examples of the use cases and features of K3D-jupyter:
- Photorealistic volume rendering for data on regular grids. Prominent example is Computer Tomography data, which can be visualized in real time with resolution of dataset 5123. Moreover K3D-jupyter experimentally supports time-series, which makes it possible to display 4D (3D + time) tomography in the browser.
- 3D point plot - high performance renderer can display millions of points in a web browser, which is suitable for e.g. point cloud data coming from 3D scanners or for real time visualization of tracers particles in fluid dynamics.
- Meshes with scalar attributes can be displayed, dynamically updated and animated.
- Voxel geometry is supported on both dense and sparse datasets. It is used for example in segmentation of CT data.
K3D-jupyter is an ipywidget, therefore it natively contains frontend and backend. Backend is a Python process where the data is prepared. Frontend is a JavaScript application with WebGL (via Three.js and custom pixelshaders) access. The ipywidget architecture allows for communication of these two parts. K3D-jupyter exposes this communication and allows for easy dataset updates on existing plot.
For example, if plt_points is an k3d.points object, then a simple assignment on the backend (Python):
plt_points.positions = np.array([[1,2,3],[3,2,1]], dtype=np.float32)will trigger data transfer of the positions to the front-end and the plot will be updated.
One of the most attractive aspects of this architecture is the fact that the backend process can run on arbitrary remote infrastructure, e.g. in the HPC center, where large scale simulations are performed or large datasets are available. Then is it possible to use the Jupyter notebook as a kind of “remote display” for visualizing that data. As the frontend is on user's computer, the interactivity of 3D inspection is very good. One can achieve fast updates on the whole dataset or any of its parts.
Similarly as in the case of matplotlib, the native data type is a numpy array. It greatly simplifies K3D usage as well as the frontend code since we implement only simple sets of objects. On the other hand, the availability of the backend does not prohibit from using much more sophisticated visualization pipelines. An example could be an unstructured volumetric grid with some scalar field. K3D does not support this kind of data, but it can be preprocessed using VTK to -for example - mesh with color coded values. Such a mesh can be an isosurface or section of the original volumetric data. Moreover, if such preprocessing produced a mesh with small or moderate number of triangles (e.g. < 105), then it could be interactively explored by linking to other widgets.
Plots can be saved as both PNG screenshots and interactive HTML snapshots, which contain the front-end JavaScript 3D application bundles with the data. They can be used independently on the Python back-end, send by email or embedded in webpages as an iframe:
Snapshots can be generated from Menu or programatically:
plot.fetch_snapshot()In the latter case one has to write HTML code to a file:
with open('../_static/points.html','w') as fp:
fp.write(plot.snapshot)