isp-uv-es/spacenet4-taco

SpaceNet 4 Off-Nadir Buildings

This is a repackaging, not a new dataset. It is SpaceNet 4 Off-Nadir Buildings by SpaceNet LLC / Maxar / CosmiQ Works, 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-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{weir2019spacenet,
  title     = {SpaceNet MVOI: A Multi-View Overhead Imagery Dataset},
  author    = {Weir, Nicholas and Lindenbaum, David and Bastidas, Alexei and Van Etten, Adam and McPherson, Sean and Shermeyer, Jacob and Kumar, Varun and Tang, Hanlin},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  pages     = {992--1001},
  year      = {2019},
  doi       = {10.1109/ICCV.2019.00108}
}

About the data

1064 ground tiles over Atlanta, each imaged by WorldView-2 at five look angles: 7 deg, 19 deg, 29 deg, 42 deg, 53 deg off nadir, spanning SpaceNet's nadir, off-nadir and very-off-nadir bands. The ground truth is one shared set of building footprints per tile, as both a mask and polygons.

1,064 samples · splits: test 213 · train 609 · validation 242 · tasks: semantic-segmentation, instance-segmentation

Packaged as TACO v3.

Full description

Layout. All five collects are orthorectified onto the same 900x900 grid at 0.5 m, so a tile is one sample holding ten rasters: the 8-band 2 m multispectral cube and the 0.5 m panchromatic band per angle. Viewing geometry is the only thing that varies across them.

Products. The released pan-sharpened products are dropped as a fusion of those two.

Empty tiles. 45 tiles carry no buildings, cross-checked against the release's own summaryData CSVs because their GeoJSON files are zero bytes rather than empty collections.

Splits. The challenge test imagery ships without labels, so there is no official split to keep. The split here is geographic, by easting: 609 train, 242 val, 213 test.

Held-out half

spacenet4_test.zip holds the 906 samples whose targets the publisher withheld. Same inputs, no target slots: it is there to be predicted on and submitted, and it belongs in no training mixture.

import os
from huggingface_hub import snapshot_download
from taco.ml import Dataset

root = snapshot_download("isp-uv-es/spacenet4-taco", repo_type="dataset",
                         allow_patterns=["spacenet4_test.zip", "spacenet4_test.zip/*", "spacenet4_test.zip/.tacocat/*"])
ds = Dataset(os.path.join(root, "spacenet4_test.zip"))

Getting started

git clone --recursive https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the reader: C++23, CMake, Ninja, pkg-config, libcurl >= 7.83, OpenSSL >= 3

Read it straight from the Hub:

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

path = hf_hub_download("isp-uv-es/spacenet4-taco", "spacenet4.zip", repo_type="dataset")
ds = Dataset(path)
plot_sample(ds[0])

or from a local copy:

ds = Dataset("spacenet4.zip")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["pan_nadir07"].array.shape

Metadata without decoding anything:

import taco
taco.read("spacenet4.zip")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	pan_nadir07	raster	optical	1 band(s), digital_number
input	ms_nadir07	raster	multispectral	8 band(s), digital_number
input	pan_nadir19	raster	optical	1 band(s), digital_number
input	ms_nadir19	raster	multispectral	8 band(s), digital_number
input	pan_nadir29	raster	optical	1 band(s), digital_number
input	ms_nadir29	raster	multispectral	8 band(s), digital_number
input	pan_nadir42	raster	optical	1 band(s), digital_number
input	ms_nadir42	raster	multispectral	8 band(s), digital_number
input	pan_nadir53	raster	optical	1 band(s), digital_number
input	ms_nadir53	raster	multispectral	8 band(s), digital_number
target	buildings	mask	label_raster	2 classes
target	footprints	polygon		unit m

Licence

CC-BY-SA-4.0

Terms of use:

• Adapted material must be shared under CC-BY-SA-4.0.

Required credits:

• The SpaceNet Dataset by SpaceNet Partners is licensed under CC-BY-SA-4.0
• Imagery: Maxar WorldView-2

Providers: SpaceNet LLC / Maxar / CosmiQ Works

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.