isp-uv-es/avi-math-taco

AVI-Math (benchmark)

This is a repackaging, not a new dataset. It is AVI-Math (benchmark) by AVI-Math authors (VisionXLab), 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:

@article{zhou2025avimath,
  author  = {Zhou, Yue and Feng, Litong and Lan, Mengcheng and Yang, Xue and Li, Qingyun and Ke, Yiping and Jiang, Xue and Zhang, Wayne},
  journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
  title   = {Multimodal Mathematical Reasoning Embedded in Aerial Vehicle Imagery: Benchmarking, Analysis, and Exploration},
  year    = {2025},
  volume  = {230},
  pages   = {289-303}
}

About the data

3773 mathematical-reasoning questions over 360 drone photographs. The questions ask for a computation on the imagery rather than a description or a location: recover the drone's above-ground level from focal length, pixel size and pitch angle, price a row of vehicles, work out the ground area a frame covers.

360 samples · splits: test 360 · tasks: visual-question-answering

Packaged as TACO v3.

Full description

Questions. Six subjects (geometry 1080, logic 767, statistics 656, arithmetic 551, counting 360, algebra 359), 20 topics and 42 named tasks, with three answer forms (free-form, multiple-choice, yes/no), a difficulty level, a solution step count and a written rationale for every question.

• The sample is the image, with its 5 to 18 questions as parallel lists.
• 1,724 questions are box-level: they refer to rectangles the release burned into a copy of the photograph, so there is no box geometry to carry, and region separates those samples from the image-level ones.
• pitch_angle and AGL are per image and constant across an image's questions.

Imagery. Consumer-drone JPEGs at 4000x2250, stored byte-for-byte, no CRS and no published transfer function, so radiometry render and no approximate scale. The sensor parameters a question needs are inside its own text, where the release put them.

Splits. test: this is an evaluation benchmark, and the release's training corpus is the separate avi-math-geomath container, whose 22,908 images do not intersect these 360.

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:

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/avi-math-taco", "avi-math.zip", repo_type="dataset")
ds = Dataset(path)
plot_sample(ds[0])

or from a local copy:

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

Metadata without decoding anything:

import taco
taco.read("avi-math.zip")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	image	raster	optical	3 band(s), render
input	questions	text_sequence	text
target	answers	text_sequence	text
target	rationales	text_sequence	text

Licence

apache-2.0

Providers: AVI-Math authors (VisionXLab)

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.