isp-uv-es/xview-taco

xView

This is a repackaging, not a new dataset. It is xView by Defense Innovation Unit Experimental / NGA (Lam et al.), Maxar WorldView-3, 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:

@article{lam2018xview,
  title   = {xView: Objects in Context in Overhead Imagery},
  author  = {Lam, Darius and Kuzma, Richard and McGee, Kevin and Dooley, Samuel and Laielli, Michael and Klaric, Matthew and Bulatov, Yaroslav and McCord, Brendan},
  journal = {arXiv preprint arXiv:1802.07856},
  year    = {2018}
}

About the data

846 WorldView-3 pan-sharpened RGB chips at 0.3 m, 2426x2912 to 3325x4199 px, with 601,774 axis-aligned objects.

846 samples · splits: test 105 · train 653 · validation 88 · tasks: object-detection

Packaged as TACO v3.

Full description

Classes. The release labels objects with a sparse code, not an index, and its own 60-entry class table does not cover the annotations.

• Codes 75 and 82 are used (51 and 28 objects) and named nowhere, so the legend here has 62 entries with those two as unlisted:75 and unlisted:82. Every object also carries its raw object_type_id.
• The authors publish two copies of the class table which disagree on five names; the one xviewdataset.org points at is used.

Annotations. Of those whose chip is on disk, 591,869 lie inside, 9,905 overhang a border and are kept with raw coordinates rather than clipped, 23 have no intersection with their chip and are dropped, and 9 are zero-area and are dropped. No chip is emptied by either drop. The geojson labels one further chip (1395.tif, 131 objects) for which no pixels are shipped, so it is not a sample. The challenge's val half is unlabelled and ships as xview_test.

Pixels. Source TIFFs are uncompressed, so re-encoding as a COG saves 39% and gives a 10 Mpx chip the internal tiling it needs.

• nodata is 0, the release's own tag: chips carry a black collar where the strip did not cover the grid cell, and zero_frac records how much of each chip it is.
• The chips are EPSG:4326, so the raw pixel size is in degrees and resolution_m is the ground value reduced at each chip's own centre latitude.
• Pixels are an 8-bit render of a WorldView product with no published stretch, so an approximate scale of 1/255 is declared and the values are not declared comparable across chips.

Overlap. ShipRSImageNet re-hosts 532 pixel-exact 920 px tiles of 114 of these chips under a different taxonomy, so the two are not independent imagery.

Held-out half

xview_test.zip holds the 281 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/xview-taco", repo_type="dataset",
                         allow_patterns=["xview_test.zip", "xview_test.zip/*", "xview_test.zip/.tacocat/*"])
ds = Dataset(os.path.join(root, "xview_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/xview-taco", "xview.zip", repo_type="dataset")
ds = Dataset(path)
plot_sample(ds[0])

or from a local copy:

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

Metadata without decoding anything:

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

Samples

What a sample contains

role	slot	holds	modality	detail
input	image	raster	optical	3 band(s), unit 1, requantised
target	boxes	bbox_2d
target	category	class_sequence		62 classes

Licence

CC-BY-NC-SA-4.0

Terms of use:

• Non-commercial use only (CC-BY-NC-SA-4.0).
• Adapted material must be shared under CC-BY-NC-SA-4.0.

Required credits:

• xView: DIUx and NGA (Lam et al.)
• Imagery: Maxar WorldView-3

Providers: Defense Innovation Unit Experimental / NGA (Lam et al.), Maxar WorldView-3

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.