Datasets:

isp-uv-es
/

gres-lisat-taco

This is a repackaging, not a new dataset. It is GRES (LISAt geospatial reasoning segmentation) by Quenum et al. (LISAt, BAIR), DIUx xView, converted to TACO. Pixel values and labels are kept as released except where the description below says otherwise. All credit belongs to the original authors: if you use it, please cite them and follow their licence.

original dataset · paper · licence: CC-BY-NC-SA-4.0

Repackaged into TACO by the Image and Signal Processing Group (ISP), Universitat de València, within the ELLIOT project.

Citation

Please cite the original work:

@inproceedings{quenum2025lisat,
  title     = {{LISAt}: Language-Instructed Segmentation Assistant for Satellite Imagery},
  author    = {Quenum, Jerome and Hsieh, Wen-Han and Wu, Tsung-Han and Gupta, Ritwik and Darrell, Trevor and Chan, David M.},
  booktitle = {Advances in Neural Information Processing Systems 38 (NeurIPS 2025)},
  pages     = {124660--124702},
  year      = {2025},
  doi       = {10.52202/085713-3756}
}

About the data

9215 512x512 aerial chips sliced from xView-1, each with one target object delineated by a polygon and described by a query that does not name it directly: reasoning segmentation rather than referring segmentation.

9,215 samples · splits: test 1,510 · train 7,205 · validation 500 · tasks: reasoning-segmentation, referring-segmentation

Packaged as TACO v3.

Full description

Queries. 23623 in total. Train chips carry three paraphrases each, val and test one.

Splits. 7205 train, 500 val, 1510 test.

Imagery. The release ships none. The pool is regenerated by slicing xView-1's 846 training chips, and the renaming is positional -- a counter over the slices in shell glob order, val before train -- so the pool is only reproducible under LC_ALL=C with exactly 846 chips. The reproduction here enumerates 36,851 slices and resolves all 9215 annotation names, 0 unmatched.

Lossy JPEG copied verbatim. No georeferencing survives the slicing, so radiometry is render and no CRS is written.

Target. Both the release's own polygon, flat and with a vertex count, and its rasterisation are stored.

Counts. The README claims 9,205 images and 27,615 queries; what ships is 9215 and 23623.

Getting started

git clone https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the native reader: needs CMake and Ninja

Read it straight from the Hub:

from huggingface_hub import snapshot_download
from taco.ml import Dataset, plot_sample

path = snapshot_download("isp-uv-es/gres-lisat-taco", repo_type="dataset")
ds = Dataset(path)                  # the repository is the container: parts + .tacocat
plot_sample(ds[0])

or from a local copy:

ds = Dataset("gres-lisat")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["image"].array.shape

Metadata without decoding anything:

import taco
taco.read("gres-lisat/.tacocat")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	`image`	raster	optical	3 band(s), render
input	`queries`	text_sequence	text
target	`mask`	mask	label_raster	2 classes
target	`polygon`	polygon

Licence

CC-BY-NC-SA-4.0

Non-commercial use only (CC-BY-NC-SA-4.0).
Adaptations must be shared under CC-BY-NC-SA-4.0.
Credit the original authors and the data sources listed below.

Required credits:

GRES / LISAt (Quenum et al.)
xView dataset (Lam et al., 2018), released by DIUx and NGA under CC BY-NC-SA 4.0

Providers: Quenum et al. (LISAt, BAIR), DIUx xView

Acknowledgements

TACO was designed by César Aybar and is specified at https://asterisk.coop/taco/spec/.

Built by Oscar Pellicer within the ELLIOT project at the Image and Signal Processing Group (ISP), Universitat de València.

Downloads last month: -