README.md

metadata

license: cc-by-nc-4.0
pretty_name: DAVIS-S
task_categories:
  - image-segmentation
tags:
  - saliency
  - salient-object-detection
  - segmentation
  - high-resolution
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

DAVIS-S

DAVIS-S is a high-resolution salient object detection (SOD) evaluation set, released alongside the HRSOD (High-Resolution Salient Object Detection) paper. It is a saliency-annotated subset of the DAVIS video object segmentation dataset, curated to benchmark high-resolution salient object detection methods. Images are full-HD (1920x1080) with pixel-accurate binary ground-truth saliency masks.

Dataset structure

• Split: test (single evaluation set), 92 image/mask pairs.
• Columns:
  • image: the RGB source image (datasets.Image).
  • mask: the grayscale ground-truth saliency mask (datasets.Image).

Usage

from datasets import load_dataset

ds = load_dataset("nobg/DAVIS-S", split="test")
ex = ds[0]
ex["image"]  # PIL.Image, RGB, 1920x1080
ex["mask"]   # PIL.Image, L (grayscale) saliency mask

Source & credits

• DAVIS dataset — the underlying images originate from the DAVIS (Densely Annotated VIdeo Segmentation) benchmark.
• HRSOD authors — the high-resolution saliency subset and ground-truth masks were released as part of the HRSOD project (yi94code/HRSOD).

Please cite the DAVIS and HRSOD works if you use this dataset.

License

Released under CC BY-NC 4.0 (non-commercial research use), consistent with the DAVIS dataset licensing.