AeroGround_v1

AeroGround_v1 is a synthetic aerial-ground reasoning benchmark with 2,250 multiple-choice questions, 50 subsets, and 11 task families. It pairs aerial and ground-view observations and provides task-specific annotations and ground-truth answers.

Scope of this release

This is the benchmark data release: the 2,250 retained questions, the 1,939 referenced original RGB observations, 186 existing candidate-image panels, and a self-contained reference input export. Source folders cover the City6 and Town1 scene groups. The original RGB observations are stored as RGBA PNGs; their RGB channels contain the color observations.

The larger simulated raw observation pool and the scene-disjoint training set described in the associated manuscript are outside this release. This package contains the benchmark split; it does not define train/validation splits. Model predictions and API evaluation scripts are maintained separately.

Task distribution

Task	Questions
distance_comparison	410
multi_target_matching	170
multi_target_relation	180
occlusion_reasoning	320
region_counting	64
relative_orientation_determination	90
reverse_sequence_reasoning	80
single_target_matching	500
trajectory_reconstruction	100
view_localization	150
view_transition_reasoning	186
Total	2,250

See metadata_summary.json for option counts, answer distributions, scene scope, and the exact sample-weighted uniform-random expectation.

Download

hf auth login
hf download ysh12321/AeroGround_v1 --repo-type dataset --local-dir AeroGround_v1

Authentication is needed when the repository is private. The original image files are downloaded by this command; Git LFS installation is not required. On Windows, use a short destination or an absolute extended-length path if temporary download paths exceed the platform path-length limit.

Use the prebuilt reference inputs

reference_inputs/benchmark_inputs.jsonl contains one record per question: prompt, ordered images, image_roles, image_sizes, option_labels, answer, and links to the detailed annotations. The 2,210 referenced JPEG assets are included. Loading these inputs needs no simulator, rendering code, API key, or network access after download.

import json
from pathlib import Path
from PIL import Image

root = Path("AeroGround_v1")
with (root / "reference_inputs/benchmark_inputs.jsonl").open(encoding="utf-8") as f:
    sample = json.loads(next(f))

model_input = {
    "prompt": sample["prompt"],
    "images": [Image.open(root / p).convert("RGB") for p in sample["images"]],
}
gold_answer = sample["answer"]  # scoring only; do not include in model_input
print(sample["sample_id"], sample["image_roles"])

This is the explicit mcq_text_reference_v1 profile, defined in reference_inputs/profile.json and REFERENCE_PROTOCOL.md. It uses coordinate text where supported and preserves all original choices, including “unable to determine” choices. It provides a reproducible reference; it does not certify that every setting matches a particular model run in the paper. Report the profile and any additional resizing or prompt changes when publishing results. Do not resize the supplied images without updating pixel coordinates in the prompt.

Index and detailed annotations

benchmark_index.jsonl is a Dataset Viewer asset inventory. image_paths_kind is available_assets: its image list is not a model input sequence. Its source_json links to the complete per-task annotations. Load the reference_inputs configuration for the prebuilt input sequence instead.

from datasets import load_dataset

inputs = load_dataset("ysh12321/AeroGround_v1", "reference_inputs", split="benchmark")
# The images column contains relative paths. Download the repository image
# assets as shown above to open those paths locally.

All downloadable asset paths use / and are relative to the repository root. Question IDs are unique within a subset; sample_id is the globally unique identifier. answer_index is zero-based. All retained annotation asset paths resolve within the release root. Unavailable historical preview/provenance path fields have been removed; geometric annotations and scoring targets are preserved. See SCHEMA.md.

Directory layout

AeroGround_v1/
  README.md
  LICENSE
  CITATION.bib
  SCHEMA.md
  REFERENCE_PROTOCOL.md
  THIRD_PARTY_NOTICES.md
  CHANGELOG.md
  metadata_summary.json
  release_manifest.json
  SHA256SUMS
  benchmark_index.jsonl
  reference_inputs/
    profile.json
    benchmark_inputs.jsonl
    images/*.jpg
  benchmark/
    Benchmark_Questions_filtered/<task>/<subset>/all_questions.json
    Benchmark_Questions/view_transition_reasoning/<subset>/visualizations/*.png
    city6/.../RGB/*.png
    town1/.../RGB/*.png

Evaluation and limitations

Compare the predicted option label with answer. Report the evaluated sample count, skipped/failed samples, input profile, model and decoding settings. Use all 2,250 questions for the full benchmark. Overall accuracy weights each sample equally; per-task accuracy and the mean of task accuracies are useful additional views. Option counts differ by task, so a single 25% chance level does not apply to every sample. Treat invalid or missing predictions as errors in the full-benchmark score, and disclose any alternative error handling.

The observations are simulated and this release covers only two source scene groups. Questions share frames and scene instances, so they are not independent real-world observations. Some tasks have uneven answer-label distributions. Ground-truth and diagnostic annotations are public for verification and must not be passed to a model as extra input. Training on these benchmark answers changes the interpretation of subsequent benchmark results.

License, attribution, and citation

The existing declared data license remains CC BY-NC 4.0; this preparation does not change it. See LICENSE and the official terms. Third-party simulation assets retain their original terms. The maintainer confirmed that the manuscript's asset-license inventory was checked against the original asset terms; attribution is recorded in THIRD_PARTY_NOTICES.md. No third-party scene source assets or engine source code are included.

Until a final paper citation is supplied, use the dataset URL and the minimal dataset citation in CITATION.bib. Do not infer an author list or DOI from a paper template. For reproducibility, record the exact Hugging Face commit or release tag used. release_manifest.json and SHA256SUMS describe this package.