EORSSD

The Extended Optical Remote Sensing Saliency Detection dataset — 2000 nadir satellite images with binary saliency masks, 1400 train / 600 test, as published.

from datasets import load_dataset

ds = load_dataset("nobg/EORSSD", split="test")   # 600 rows
ds[0]["image"]       # PIL, the original JPEG
ds[0]["mask"]        # PIL, the binary label PNG
ds[0]["image_name"]  # '0004' — the numeric stem the two are joined on

Why this mirror exists

EORSSD is the cheapest way to add a genuinely different imaging geometry to a background-removal or salient-object benchmark: nadir view, no horizon, no perspective size cue, and objects that are often a few dozen pixels across. Sets built from ground-level photography — DIS5K, DUTS, HRSOD, COD10K and the rest — all share a manifold that this one does not.

The authors distribute it as a single 63 MB zip with a flat four-folder layout. This mirror is that zip in parquet, joined on the numeric filename stem, with image and label bytes passed through unmodified — verified by SHA-256 on all 4000 members. Nothing is decoded or re-encoded, which matters for the labels specifically: re-encoding a binary mask can introduce intermediate grey values and quietly change what every downstream metric measures.

Published comparator: SAM2-UNeXT reports S_α 0.948 on the 600-image test split.

Two upstream properties to know before you score it

70 masks are entirely black — 52 in train, 18 in test (3.0 % of the test split). These are EORSSD's deliberate no salient object scenes, not corrupt rows, and they are kept so the split stays the published 600. But they interact badly with the usual metrics: with empty ground truth, S_α and MAE still behave sensibly, while IoU, boundary-IoU and F-max are 0 for any non-empty prediction. A model whose output is never empty by construction — most matting models — is pinned at 0 on those three metrics for these rows, so quote S_α/MAE here and treat the rest with care.

The test labels are not uniformly 8-bit binary. 595 are mode L, 4 are RGB and 1 is RGBA, and 5 carry anti-aliased edges (142–237 grey levels rather than 2). Call .convert("L") and binarize rather than assuming a single-channel two-level PNG. Nothing here re-encodes the labels, so this is exactly what the authors shipped.

Licensing

No license is declared by the authors (the GitHub API reports license: null). It is left unset here rather than guessed; check with the original authors before any use beyond research.

Citation

@article{zhang2020dense,
  title={Dense Attention Fluid Network for Salient Object Detection in Optical Remote Sensing Images},
  author={Zhang, Qijian and Cong, Runmin and Li, Chongyi and Cheng, Ming-Ming and Fang, Yuming and Cao, Xiaochun and Zhao, Yao and Kwong, Sam},
  journal={IEEE Transactions on Image Processing},
  volume={30},
  pages={1305--1317},
  year={2021}
}