including optical flow, occlusions, correspondences, segmentation labels, normals, and depth
Creative Flow+ Dataset challenges Computer Vision techniques to generalize to a wide range of styles, including messy stylized content. Our dataset is the first diverse multi-style artistic video dataset densely labeled with ground truth. Our synthetic dataset contains 3000 animated sequences (124K train and 10K test frames at 1500x1500) rendered in a wide range of artistic styles (40 line styles and 38 shading styles).
@InProceedings{shugrina2019creative,
author = {Shugrina, Maria and
Liang, Ziheng and Kar, Amlan and Li, Jiaman and
Singh, Angad and Singh, Karan and Fidler, Sanja},
title = {Creative Flow+ Dataset},
booktitle = {The IEEE Conference on Computer Vision
and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}
Our full dataset is provided in multiple compressed packages at full (1500x1500)
and half (750x750) resolution.
IMPORTANT: To avoid surprises, please read the first section of our errata. You may also sign up to be notified of any changes.
IMPORTANT: Go to our github repository for instructions and scripts to decompress these downloads.
IMPORTANT: The links on this webpage are invalid until you click one of them to accept the user agreement.
Download these lists for navigating the data with our python utilities.
These packages include forward flow, as special zip files, and occlusions, as mp4 clips.
These packages include core metadata, compression strategies vary (see errata).
The _supp_ packages include raw normals and correspondences images, compressed losslessly.
The _depth_ packages include raw depth arrays, compressed losslessly.
The _backflow_ packages include back flow from frame F(i) to frame F(i-1).
Data sanity packages include per-frame sanity computed for full resolution frames.
See errata for details.
Below are the named shading and line styles. Our python utilities can easily filter sequences by style name, if you wish to only train/test on a subset. Below we also include videos of the training styles to give a sense of the styling in action, in the presence of color randomization.
Data Pipeline
Our data generation pipeline is available on our github page.