SpatialUAV

Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion

SpatialUAV is a diagnostic benchmark for evaluating spatial intelligence in real low-altitude unmanned aerial vehicle (UAV) scenarios. It contains 4,331 curated visual question-answering instances across 14 task types, covering semantic discrimination, spatial relations, aerial-aerial collaboration, aerial-ground collaboration, and motion understanding.

The benchmark provides seven visual-input configurations and nine answer formats under a unified visual-input-question-answer schema. It is intended for benchmark evaluation and contains a single test split.

Dataset Configurations

Configuration	Split	Rows	Purpose
full	test	4,331	Full benchmark evaluation
subset_20pct_per_task	test	863	Faster development and human-reference subset

Load the annotation tables with datasets:

from datasets import load_dataset

full = load_dataset("Hyu-Zhang/SpatialUAV", "full", split="test")
subset = load_dataset(
    "Hyu-Zhang/SpatialUAV",
    "subset_20pct_per_task",
    split="test",
)

The Parquet files contain image paths rather than embedded image bytes. Download and extract the visual archives before resolving these paths.

Record Structure

Each annotation contains five fields:

Field	Type	Description
id	string	Unique identifier; its prefix denotes the task
image	list[string]	Ordered paths to images or video frames
conversations	list[object]	Question or instruction in {from, value} format
source	string	Dataset source tag (SpatialUAV)
GT	string	Canonical ground-truth answer

Example:

{
  "id": "Region_Recognition_00001",
  "image": ["./SpatialUAV/samples_Single_Image/img0001.jpg"],
  "conversations": [
    {
      "from": "human",
      "value": "Which regions in the image contain a parking lot? Answer with only the region labels, formatted exactly like `Region 1, 2`. No explanation."
    }
  ],
  "source": "SpatialUAV",
  "GT": "Region 3, 4."
}

The full annotations are provided as annotations.jsonl, annotations_shuffled.jsonl, and the viewer-friendly annotations_shuffled.parquet.

Downloading and Extracting the Visual Data

Clone the dataset repository with Git LFS:

git lfs install
git clone https://huggingface.co/datasets/Hyu-Zhang/SpatialUAV
cd SpatialUAV

The aerial-ground archive is distributed in multiple parts. Reassemble and verify it before extraction:

cat A2G.zip.part-* > A2G.zip
sha256sum -c A2G.zip.sha256  # Linux
# shasum -a 256 -c A2G.zip.sha256  # macOS

Extract the visual archives in the repository root:

unzip A2A.zip
unzip A2G.zip
unzip samples_Single_Image.zip
unzip samples_Motion_Understanding_Frames.zip

The resulting directories are:

SpatialUAV/
├── samples_Single_Image/
├── samples_A2A_Pured/
├── samples_A2A_detected/
├── samples_A2A_Occlusion_Removal/
├── samples_A2G_Pured/
├── samples_A2G_detected/
├── samples_A2G_Path_Planning/
└── samples_Motion_Understanding_Frames/

Annotation paths begin with ./SpatialUAV/ and resolve when evaluation is launched from the parent directory of the cloned repository.

More Information

For task definitions, dataset construction, evaluation metrics, model inference, and benchmark results, see the GitHub repository and the SpatialUAV paper.

Licensing

The code in the SpatialUAV GitHub repository is released under the MIT License. The benchmark data is not relicensed under MIT. It is derived from multiple source datasets and remains subject to their respective licenses and terms of use. Accordingly, this Hugging Face dataset is marked with license: other.

Users are responsible for reviewing and complying with the terms of BEDI, AirCopBench, MAVREC, AirScape, University-1652, and any upstream resources incorporated by those datasets.

Citation

If you use SpatialUAV, please cite:

@article{zhang2026spatialuav,
  title   = {SpatialUAV: Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion},
  author  = {Zhang, Haoyu and Liu, Meng and Xiang, Qianlong and Wang, Kun and Wang, Yaowei and Nie, Liqiang},
  journal = {arXiv preprint arXiv:2606.27876},
  year    = {2026}
}

Please also cite the applicable source datasets when using the corresponding visual data.