Datasetpublic

emirkisa/DAVIS-2017-480p-mp4

DAVIS 2017 — 480p MP4 Sequences Pre-encoded MP4 versions of all 90 sequences from the DAVIS 2017 trainval split at 480p resolution. Useful if you want to quickly load, stream, or embed DAVIS sequences without running ffmpeg yourself or dealing with per-frame JPEG folders. Contents File pattern Description <seq>_raw_24fps.mp4 Raw RGB frames, no annotation, 24 fps <seq>_ov055_24fps.mp4 RGB + DAVIS palette mask overlay at 55% opacity, 24 fps… See the full description on the dataset page: https://huggingface.co/datasets/emirkisa/DAVIS-2017-480p-mp4.

sourceHugging Facecc-by-nc-4.0updated 6mo agoView on Hugging Face

1likes1kdownloads

Dataset Card

DAVIS 2017 — 480p MP4 Sequences

Pre-encoded MP4 versions of all 90 sequences from the DAVIS 2017 trainval split at 480p resolution.

Useful if you want to quickly load, stream, or embed DAVIS sequences without running ffmpeg yourself or dealing with per-frame JPEG folders.

File pattern	Description
`<seq>_raw_24fps.mp4`	Raw RGB frames, no annotation, 24 fps
`<seq>_ov055_24fps.mp4`	RGB + DAVIS palette mask overlay at 55% opacity, 24 fps
`<seq>_ov010_24fps.mp4`	RGB + DAVIS palette mask overlay at 10% opacity (subtle), 24 fps

—90 sequences × up to 3 variants = 185 MP4 files
—Total size: ~290 MB
—Codec: H.264 (libx264, yuv420p, CRF 18, faststart)
—Source annotations: per-pixel palette-indexed PNGs where pixel value = object ID

Usage

python

from huggingface_hub import hf_hub_download

# Download one sequence (raw)
path = hf_hub_download(
    repo_id="emirkisa/DAVIS-2017-480p-mp4",
    filename="camel_raw_24fps.mp4",
    repo_type="dataset",
)

# Download everything (~290 MB)
from huggingface_hub import snapshot_download
local_dir = snapshot_download(
    repo_id="emirkisa/DAVIS-2017-480p-mp4",
    repo_type="dataset",
)

DAVIS Colour Palette (overlay videos)

Objects are coloured with the official DAVIS 20-colour palette: object 1 → #800000, object 2 → #008000, object 3 → #808000, … Background pixels are fully transparent (not blended).

Splits

The 90 sequences span both DAVIS-2016 and DAVIS-2017:

Split	Sequences
DAVIS-2016 train	30
DAVIS-2016 val	20
DAVIS-2017 train	60
DAVIS-2017 val	30

(Sequences appear in multiple splits; the total unique count is 90.)

License

Inherited from the original DAVIS dataset: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) — free for non-commercial use with attribution.

Citation

bibtex

@article{Pont-Tuset_arXiv_2017,
  author  = {Jordi Pont-Tuset and Federico Perazzi and Sergi Caelles and
             Pablo Arbelàez and Alexander Sorkine-Hornung and Luc Van Gool},
  title   = {The 2017 {DAVIS} Challenge on Video Object Segmentation},
  journal = {arXiv:1704.00675},
  year    = {2017}
}