Dynamic Flash Video Data (Ready-to-use) with Time-varying Appearance

A dataset of 21 flash-lit videos of time-varying materials is acquired, by a simple-to-construct setup. The data is pre-processed by the following protocol, and ready to use: each video has 100 JPG frames that are equally-spaced sampled from raw data, with the spatial resolution of 4032 x 3024. The data includes many natural phenomena: rusting, melting, drying, crystallizing, etc.

The synthetic data is created by Blender. The details of physical capture setup can be found in our paper, and some instructions to reproduce this dataset can be found in this repo.

This dataset is released in the paper:
Chen Liu and Tobias Ritschel. 2024. Neural Differential Appearance Equations. ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia) (2024).