Note

This is a community plugin, an external project maintained by its respective author. Community plugins are not part of FiftyOne core and may change independently. Please review each plugin’s documentation and license before use.

davis-2017#

A FiftyOne remote zoo dataset for DAVIS 2017 — the Densely Annotated VIdeo Segmentation challenge release used for the semi-supervised video object segmentation track. Data are organized as per-frame samples (or as generated video clips) with instance-level masks and boxes. Official challenge hub: davischallenge.org/davis2017/code.html.

Source, citation, and license#

Challenge / download page: https://davischallenge.org/davis2017/code.html
Evaluation code (reference implementation): davisvideochallenge/davis2017-evaluation
License: same terms as the DAVIS release — see LICENSE in davis2017-evaluation
Citation:

@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}
}

Quick start#

Install FiftyOne, then load the dataset by GitHub URL (downloads and caches automatically when needed):

pip install fiftyone

import fiftyone as fo
import fiftyone.zoo as foz

dataset = foz.load_zoo_dataset(
    "https://github.com/voxel51/davis-2017",
    split="validation",  # optional, for a specific split
    max_samples=500,  # optional, for quick exploration
    format="image",  # "image" (default) or "video"
)

session = fo.launch_app(dataset)

# Dynamically Grouped Dataset
grouped_view = dataset.group_by("sequence_id", order_by="frame_number")

# Explore a single sequence
seq_view = dataset.match_tags("blackswan")

Notes:

Supported split values: train, validation, test-dev, test-challenge.
Resolution: 480p (matches the DAVIS-2017-*-480p.zip archives fetched from the official host)
Image mode creates one sample per frame. Video mode stitches each sequence into an .mp4 and creates a local copy under the dataset tree. Attaches ground_truth on frames where annotations exist.

Statistics#

Split	Sequences	Total Samples	Annotated Samples
train	59	4,209	4,209
validation	30	1,999	1,999
test-dev	30	2,086	30
test-challenge	30	2,180	30

Training and validation splits are fully annotated. For test-dev and test-challenge, only the first frames carry dense masks (hence the small “annotated” counts in the table); remaining frames appear without ground_truth detections.

Sample fields#

Each sample includes:

filepath — JPEG for that frame (image / group modes) or generated video path (format="video").
tags —
- DAVIS subset name (train, val, test-dev, or test-challenge; note the zoo split is validation but the tag is val)
- DAVIS the sequence name
sequence_id — DAVIS sequence (clip) name. Note: Sequences do not overlap accross splits.
frame_number — frame id within the sequence (image / group modes; frame index in video mode).
ground_truth — fo.Detections with one Detection per object: normalized bounding box, instance mask (mask), and label string {sequence_id}{object_index}.

Visualization#

Example grid with the dataset loaded in image format:

DAVIS Sample Visualization