isp-uv-es/satin-aid-multilabel-taco
AID MultiLabel (SATIN mirror) 3,000 aerial scenes at 600x600 from AID, relabelled with the 17 object and cover categories present in each -- airplane, cars, dock, mobile home, ship, tanks -- rather than with the one scene class AID itself assigns. The legend is the same 17 as dlrsd's, over different pixels and at 600 px instead of 256, so the two are a matched pair for asking whether a multi-label head generalises across resolution and source. 3,000… See the full description on the dataset page: https://huggingface.co/datasets/isp-uv-es/satin-aid-multilabel-taco.
<p align="center"><img src="assets/isp.png" width="173" alt="Image Processing Laboratory"> <img src="assets/elliot.png" width="313" alt="Elliot"></p>
AID MultiLabel (SATIN mirror)
3,000 aerial scenes at 600x600 from AID, relabelled with the 17 object and cover categories present in each -- airplane, cars, dock, mobile home, ship, tanks -- rather than with the one scene class AID itself assigns. The legend is the same 17 as dlrsd's, over different pixels and at 600 px instead of 256, so the two are a matched pair for asking whether a multi-label head generalises across resolution and source.
3,000 samples · splits: test 600 · train 2,100 · validation 300 · tasks: scene-classification
Packaged as TACO v3.
<details> <summary>Full description</summary>
Source. Fetched as the SATIN benchmark's Hugging Face mirror, which ships one Parquet with the image bytes inline. JPEG rows are stored byte-for-byte; PNG and TIFF rows are re-encoded as COGs.
Splits. The release publishes none, so the split here is constructed, 70/10/20, and marked as such in split_source.
Georeferencing. No CRS and no geotransform. These are 8-bit renders, declared render with no approximate reflectance mapping, because there is no reflectance axis to approximate.
</details>
Getting started
git clone https://github.com/OscarPellicer/taco
pip install -e "taco[ml]"Read it straight from the Hub:
from huggingface_hub import hf_hub_download
from taco.ml import Dataset, plot_sample
path = hf_hub_download("isp-uv-es/satin-aid-multilabel-taco", "satin-aid-multilabel.tacozip", repo_type="dataset")
ds = Dataset(path)
plot_sample(ds[0])or from a local copy:
ds = Dataset("satin-aid-multilabel.tacozip")
sample = ds[0] # {slot name: SlotValue}, arrays decoded
sample["image"].array.shapeMetadata without decoding anything:
import taco
taco.read("satin-aid-multilabel.tacozip") # one Arrow table, levels joinedSamples
What a sample contains
Licence
unspecified (AID states no terms; SATIN itself says only "can be used for research purposes")
Providers: Xia, Hu et al. (Wuhan University) -- imagery, Hua, Mou, Zhu (DLR / TU Munich) -- multi-labels, SATIN benchmark (Roberts et al., 2023)
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 Processing Laboratory, Universitat de València.
