shijli/davis2017

DAVIS 2017

DAVIS 2017 (Densely Annotated VIdeo Segmentation), the semi-supervised video object segmentation benchmark, packed once from the official 480p trainval archive with one row per frame, so it loads in one line and no data path has to be configured:

from datasets import load_dataset

ds = load_dataset("shijli/davis2017", "semi-supervised")  # 60 train / 30 val videos, 480p

In semi-supervised video object segmentation, the mask of the first frame is given and the objects have to be segmented in every following frame. DAVIS 2017 has several objects per video and is scored at 480p with the region similarity J, the boundary accuracy F and their mean J&F.

Splits

split	videos	frames	objects	official list
train	60	4209	144	ImageSets/2017/train.txt
validation	30	1999	61	ImageSets/2017/val.txt

Only trainval is packed: the test-dev and test-challenge sets ship first-frame masks only, so they cannot be scored offline.

Columns

column	type	content
id	string	<video>/<frame file stem>, e.g. bear/00000: the official file names, so rows join to a local DAVIS folder
video	string	video name
frame	int32	0-based index of the frame in its video
image	image	the official 480p JPEG
mask	image	the official annotation: palette PNG whose pixel value is the object id in the video, 0 background, 255 void
num_objects	int32	number of objects of the video, the largest id of its first-frame mask

Object ids are per video, not classes: object 1 of bear has nothing to do with object 1 of dogs-jump. The rows of a video are contiguous and in frame order, so a video is a groupby:

import itertools

val = load_dataset("shijli/davis2017", "semi-supervised", split="validation")
for video, frames in itertools.groupby(val, key=lambda row: row["video"]):
    frames = list(frames)  # frames[0]["mask"] is the given first-frame mask

How this was packed

The repository includes create_dataset.py, which produced these parquet files from the official DAVIS-2017-trainval-480p.zip. Images and masks are copied through byte for byte. Before pushing, the build was checked against the source: every split holds the videos and frames of its official list, the rows of every video are contiguous and in frame order, and sampled masks are mode P, hold only the video's object ids or void, match their image's size and are byte-identical to the source files.

License and attribution

DAVIS 2017 is released by its authors under CC BY-NC 4.0. The videos come from the online sources listed in SOURCES.md, which keep their own terms of use. This packaging adds no rights: use it under those same terms and cite the original work.

@article{Pont-Tuset_arXiv_2017,
    author = {Jordi Pont-Tuset and Federico Perazzi and Sergi Caelles and Pablo Arbel\'aez and
              Alexander Sorkine-Hornung and Luc {Van Gool}},
    title = {The 2017 DAVIS Challenge on Video Object Segmentation},
    journal = {arXiv:1704.00675},
    year = {2017}
}
@inproceedings{Perazzi_CVPR_2016,
    author = {Federico Perazzi and Jordi Pont-Tuset and Brian McWilliams and Luc {Van Gool} and
              Markus Gross and Alexander Sorkine-Hornung},
    title = {A Benchmark Dataset and Evaluation Methodology for Video Object Segmentation},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2016}
}