AerialPerson-Mix

Single-class (person) YOLO labels for 3,985 aerial images from SARD, HERIDAL, LADD and NOMAD. This is the training and validation data for the AerialPerson-YOLO26 detector, described in the IEEE SSRR 2026 paper One Operator, Many Robots: Natural-Language Command of UAV–UGV Teams for Search, Victim Detection, and Rescue.

• Archive and DOI: https://doi.org/10.5281/zenodo.20643213 (always resolves to the latest version)
• Models: https://huggingface.co/alejp1998/AerialPerson-YOLO26

Exact single-class (person) YOLO labels, split membership, and a verified rebuild procedure for the paper's four-source training set. Original images are not redistributed; reconstruct_aerialperson_mix.py downloads them from their providers and checks every image before use.

Source	Training images	Validation images	Total
HERIDAL	788	197	985
LADD	800	200	1,000
NOMAD	800	200	1,000
SARD	800	200	1,000
Total	3,188	797	3,985

Contents

• labels/train/, labels/val/: one YOLO label file per manifest image; validation contains 2,188 person boxes.
• manifests/all_used_images.csv: source, split, bundle filename and original identity of every image; per-source CSVs, train.txt/val.txt and counts.csv agree with it.
• reconstruct_aerialperson_mix.py: downloads the provider exports and rebuilds images/ beside these labels (standard library only, plus kaggle for LADD).
• IMAGE_SHA256SUMS.txt: SHA-256 of all 3,985 images as rebuilt on 9 October 2026 — the images used for the revision's source-specialist sweep.
• reconstruction_report.json: verification summary of that rebuild.

Rebuild the images

Requirements: Python 3.10+, about 21 GB for downloads and 15 GB for the result, a free Roboflow API key (SARD and HERIDAL) and a Kaggle API token saved as ~/.kaggle/kaggle.json (LADD). NOMAD is public. From this folder:

python3 -m pip install kaggle
export ROBOFLOW_API_KEY=<your key>
python3 reconstruct_aerialperson_mix.py download --sources sources
python3 reconstruct_aerialperson_mix.py build --sources sources     --output AerialPerson-Mix --expected-sha256 IMAGE_SHA256SUMS.txt

download resumes and accepts --only sard,heridal,ladd,nomad. It fetches:

Source	Provider export	Fetched	Size
SARD	Roboflow rtiaditya-yoga-darmawan/sar-detection v9, YOLOv11 format (provider-resized to 832×832)	full export, 1,980 images	0.4 GB
HERIDAL	Roboflow new-workspace-qsuag/heridal-t428z v2 ("yolo_non_412"), TensorFlow CSV format, no preprocessing	full export, 1,546 images	4.1 GB
LADD	Kaggle maxinstellar/lost-people-detection (COCO and VOC annotations)	full dataset	9.2 GB
NOMAD	NOMAD Google Drive folder (images/ and labels/)	only the 1,000 listed frames and their labels	6.6 GB

build then (1) locates every manifest image under its provider name; (2) converts the provider's own annotation exactly as the historical builder did and requires it to equal the retained label: SARD pose classes merged to person; HERIDAL CSV boxes; LADD COCO boxes not flagged ignore; NOMAD YOLO boxes after the historical size filter, which kept boxes whose shorter side was at least 40/20/15 px at 10/30/50 m in the native 5472×3078 frames; (3) copies each image under its bundle filename; and (4) with --expected-sha256, requires byte-identical images. Any failure stops the build before an output is written; an existing output directory is refused. The result contains images/, these labels/, data.yaml (path: .), IMAGE_SHA256SUMS.txt and reconstruction_report.json, and can be passed directly to Ultralytics.

The October 2026 rebuild matched all 3,985 retained labels. A fresh SARD export was byte-identical to the February 2025 export used originally. Re-evaluating the released checkpoints on the rebuilt validation images reproduced all 45 previously recorded results exactly (precision, recall, mAP50, mAP50–95; see A_detector/source_training_comparison/reconstruction_fidelity.csv). If a provider later re-encodes an export, the checksum step reports the changed files while the annotation step still confirms image identity.

Scope

NOMAD filenames preserve actor, distance stratum and frame. The 10/30/50 m strata are retained, and the same 88 actors occur in both training and validation; this is not an actor-held-out test. Manifests do not provide nadir/oblique viewpoint labels. This is a curated validation split used for development/selection, not an independent field test or background-only false-alert evaluation. The historical per-source selection preferred images with larger boxes; the manifests record its outcome, so rebuilding does not rerun that selection.

Converted labels/definition use the existing CC-BY-4.0 terms. Obtaining and using original imagery remains subject to each original provider's terms.

Detector results on this data

Validation mAP50–95 of YOLO26l trained on all four sources or on one source. The detector trained on this dataset is AerialPerson-YOLO26.

How to cite

If you use these data or models, please cite the paper and this record.

@inproceedings{jaraboPenas2026OneOperator,
  author    = {Jarabo-Pe{\~n}as, Alejandro and
               Bravo-Arrabal, Juan and
               Ser{\'o}n-Barba, Javier and
               Lin-Yang, Da-hui and
               Pastor, Francisco and
               Ladig, Robert and
               Fern{\'a}ndez-Lozano, Juan Jes{\'u}s and
               Garc{\'i}a-Cerezo, Alfonso and
               Christensen, Anders Lyhne},
  title     = {One Operator, Many Robots: Natural-Language Command of
               {UAV}--{UGV} Teams for Search, Victim Detection, and Rescue},
  booktitle = {2026 IEEE International Symposium on Safety, Security,
               and Rescue Robotics (SSRR)},
  year      = {2026},
  publisher = {IEEE},
  note      = {In press}
}

Record: https://doi.org/10.5281/zenodo.20643213 (concept DOI, always resolves to the latest version)