Animating Pictures with Eulerian Motion Fields

Dataset page

This page contains information and downloads for the fluid motion stock-footage dataset used in the paper.

Dataset contents

The dataset consists of short video clips extracted from longer stock-footage videos. All files are prepended with two numbers that identify the video clip (#####_#####): the first number is the source video ID, the second number is the frame offset within that video. For clips with the same source video ID, the contents are likely to be very similar, but note that they are not identical --- the video clip, input image, and motion fields are taken from different sections of the source video.

Files in the dataset include:

#####_#####_gt.mp4: The ground-truth video clip. All videos are 60 frames in duration, and have 1280 px width and variable height. Encoded as H264 MP4s with excessively high bitrates to avoid compression artifacts. Code to load videos as PyTorch tensors can be found in here.
#####_#####_input.jpg: The first frame of each video, for convenience. Used for training the motion estimation network. Images are the same size as the corresponding video.
#####_#####_motion.pth: The "ground-truth" motion fields, which were estimated from the video using off-the-shelf optical flow networks. Stored as compressed binaries, code for reading these files can be found here.

Downloads

	Contents	Subset	Format	Size	Links
Motion estimation pairs	`[...]_input.jpg` `[...]_motion.pth`	Training: Validation:	.zip .zip	13.2GB 253MB	(Google Drive) (Google Drive)
Warping clips	`[...]_gt.mp4` `[...]_motion.pth`	Training: Validation:	.zip .zip	41.2GB 1.1GB	(Google Drive) (Google Drive)
Testing images	`[...].png`	Test:	.zip	35MB	(Google Drive)
Everything	`[...]_input.jpg` `[...]_motion.pth` `[...]_gt.mp4` `[...].png`	Train, Test, Validation:	.zip	42.9GB	(Google Drive)

Terms of Use

The contents of this dataset are licensed under the Creative Commons Attribution 4.0 License. The contents of the dataset were derived from videos downloaded from Storyblocks.com, under the following license agreement with the University of Washington.

If you use this dataset for any commercial or research purposes, please include the below citation to our work.

BibTeX

@InProceedings{Holynski_2021_CVPR,
    author    = {Holynski, Aleksander and Curless, Brian L. and Seitz, Steven M. and Szeliski, Richard},
    title     = {Animating Pictures With Eulerian Motion Fields},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {5810-5819}
}