isp-uv-es/dvl-instruct-taco

DVL-Instruct (image sets)

This is a repackaging, not a new dataset. It is DVL-Instruct (image sets) by DynamicVL authors (Xuan et al.), USDA NAIP (imagery, US public domain), 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: apache-2.0

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{xuan2025dynamicvl,
  title     = {DynamicVL: Benchmarking Multimodal Large Language Models for Dynamic City Understanding},
  author    = {Xuan, Weihao and Wang, Junjue and Qi, Heli and Chen, Zihang and Zheng, Zhuo and Zhong, Yanfei and Xia, Junshi and Yokoya, Naoto},
  booktitle = {Advances in Neural Information Processing Systems 38 (NeurIPS 2025) Datasets and Benchmarks Track},
  year      = {2025},
  eprint    = {2505.21076},
  archivePrefix = {arXiv}
}

About the data

2193 multi-temporal NAIP image sets over 42 US cities, 2005-2023, at 1.0 m and 1024x1024 px with 4-10 frames each. This is the image-set half of DVL-Suite (DynamicVL), whose tasks are basic change analysis, change-speed estimation, referring change detection and dense temporal captioning.

2,193 samples · splits: test 500 · train 1,524 · validation 169 · tasks: change-detection, referring-segmentation, visual-question-answering, change-captioning

Packaged as TACO v3.

Full description

Targets. The release's semantic change masks, which encode a transition as 10*from + to over five classes. All 21 occurring codes are named.

• The legend is shipped in no file. It is confirmed three independent ways: the authors' own code, the referring-mask filenames which spell the pair in words, and the pixel spectra, which separate vegetation from bare ground by 70 units of NIR-minus-red.
• Per set the release gives one mask per adjacent year pair plus one for the whole span. The adjacent ones form a mask_series in time order; the endpoint mask is separate, because over an endpoint span a pixel can change twice and the two are different labellings.
• Referring masks are carried where an expression references them. The release ships many more masks than expressions (93.8% of the test masks and 38.1% of train are unreferenced); n_refer_orphan_masks records the rest, which cannot be aligned to a text sequence and are not reconstructible from the semantic masks, the referring masks being dilated.

Near-infrared. The NIR is a separate leaf, a property of NAIP rather than a storage convenience. 663 of the 14,871 frames have no NIR at all, and they are exactly the pre-2010 acquisitions: 100% of 2005 and 2006, 46% and 72% through the 2008-2009 transition, and none of the 12,993 frames from 2010 to 2023, NAIP having added the band progressively until it became standard around 2010.

• The band was never acquired for those frames, so there is nothing to recover.
• It is not a padded fourth channel, because uint8 has no sentinel left: 36% of observed NIR planes contain 0 and 22% contain 255, so either choice would mask real pixels.
• rgb/ carries every frame at three bands, nir/ only the frames that have one, and nir_frame_index plus has_nir give the exact alignment.
• Splitting the container by band count is impossible: 567 of the 2,193 sets mix the two, and no set is uniformly three-band.

Encoding. Source TIFFs and PNGs are re-encoded as COGs, losslessly and smaller than the release's own files (0.814x for the imagery, 0.486x for the semantic masks). The release carries no georeferencing, so no CRS and no transform are written and the stated 1.0 m GSD lives on the bands.

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/dvl-instruct-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("dvl-instruct")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["rgb"].array.shape

Metadata without decoding anything:

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

Samples

What a sample contains

role	slot	holds	modality	detail
input	rgb	raster_series	optical	3 band(s), unit 1, requantised
input	nir	raster_series	optical	1 band(s), unit 1, requantised
input	refer_expressions	text_sequence	text
target	cd_adjacent	mask_series	label_raster	56 classes
target	cd_endpoint	mask	label_raster	56 classes
target	refer_masks	mask_set	label_raster
target	dense_temporal_caption	text_sequence	text

Licence

apache-2.0

Providers: DynamicVL authors (Xuan et al.), USDA NAIP (imagery, US public domain)

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.