This repository contains supplementary material for the paper Utilizing Deep Learning-Based Video Frame Interpolation for Scientific Visualization.
We provide the videos corresponding to the keyframe selections presented in Table 1 of the paper (table1_*.mp4), as well as side-by-side comparisons of the ground truth, FILM and RIFE when retaining only every second, fourth, and eighth frame as a keyframe (time_stride(2|4|8)_GT_FILM_RIFE.mp4).
The provided prototype consists of a Catalyst 2.0 adapter (adapter.py) and a pipeline file (pipeline.py).
The prototype illustrates the integration of ParaView Catalyst rendering, PyCinema frame interpolation via FILM, and MS-SSIM-based image quality evaluation using LIBRA for adaptive keyframe selection.
The code is intended primarily as a starting point for understanding how the individual components of the workflow are connected. It is not intended as a standalone or turnkey reproducibility package. Execution requires an environment featuring:
- ParaView/Catalyst: https://gitlab.kitware.com/paraview/paraview
- PyCinema: https://github.com/cinemascience/pycinema
- LIBRA: https://github.com/lanl/libra
- FILM VFI: https://github.com/google-research/frame-interpolation
Setting up these dependencies is outside the scope of this supplementary material. The scripts assume that the required packages can be imported from the Python environment in which the Catalyst pipeline is executed and that the FILM model is available at the configured model path.
The rendering parameters contained in the pipeline, including the camera configuration, transfer functions, scalar field, and volume-rendering settings, correspond to the particular visualization used in this example. They can be adapted to other Catalyst pipelines and datasets.