Aerial Person Detection Test Set

Aerial Person Test Dataset Description Aerial Person Test Dataset is a held-out dataset developed for evaluating person detection models in aerial, drone, webcam, and top-view imagery. The complete dataset used in the study comprises 500 test images and 50,722 annotated instances of the person class. It was used exclusively for external evaluation and was not included in model training, validation, hyperparameter tuning, or model selection. The public Zenodo record is metadata-only. It documents the dataset used in the experiments but does not contain original videos, extracted image frames, or YOLO annotation files. Classes The dataset contains a single class: 0 - person The annotations used internally follow the YOLO object detection format: class_id x_center y_center width height All bounding-box coordinates are normalized to the interval [0, 1]. Data collection and annotation The visual material was collected from several web-based sources, including stock-video platforms, public video platforms, and public webcam or live-stream services. The documented sources include Pexels, Vecteezy, YouTube, Webcam Romania, SkylineWebcams, Bouwwebcam, and See Transylvania, where applicable. The selected material covers several scenarios relevant to aerial person detection, including high-altitude views, top- view and oblique perspectives, small-scale persons, variable crowd densities, urban environments, open areas, and different illumination conditions. The images were initially pre-annotated using a YOLO-based person detection pipeline. All generated annotations were subsequently reviewed manually in Label Studio. Missing persons were annotated, inaccurate bounding boxes were corrected, and false detections were removed before the annotations were exported in YOLO format. Source provenance Detailed source-provenance information is provided in the accompanying SOURCES and source-provenance files. For each identified source, the documentation includes: • the original platform and source URL; • the corresponding source video or camera; • the identifiers of the derived dataset images; • extraction timestamps or frame ordinals, where available; • the applicable license or platform terms; • the date on which the licensing information was accessed; • attribution requirements, where applicable; • the decision regarding inclusion or exclusion of the original frames from public redistribution. Repository contents The public Zenodo deposit provides: • dataset documentation; • source-provenance documentation; • image-to-source mappings; • extraction timestamps or frame ordinals, where available; • dataset configuration information; • licensing and reuse notes; • privacy and ethics information; • traceability information. This public Zenodo record is metadata-only. It does not contain original videos, extracted image frames, or YOLO annotation files. The complete image and annotation set used during the experiments remains separate from the publicly accessible documentation because redistribution rights could not be demonstrated for all web-sourced materials. Dataset split Test set: 500 images No training or validation split is provided because the dataset was designed exclusively as an external held-out test set. Access and reuse The files in this public metadata record are openly accessible. However, the public record does not include the original image frames, videos, or YOLO annotation files. The original visual materials remain subject to the copyright conditions, licenses, and terms of use established by their respective creators, platforms, and stream providers. The presence of a source URL in the documentation does not constitute authorization to download, extract, reproduce, or redistribute the corresponding content. Where redistribution rights for extracted frames could not be demonstrated, the original frames are excluded from the public deposit. Instead, the record provides source references, dataset image identifiers, extraction metadata, licensing notes, and traceability information. Any researcher wishing to reconstruct or reuse part of the dataset must independently verify the current license, copyright, privacy, and platform requirements associated with each original source. Redistribution of the original image frames, videos, annotations, metadata, or derived dataset components is not permitted without prior approval from the record owners and without compliance with the terms of the original sources. Further information is provided in the SOURCES, RIGHTS, TRACEABILITY, ETHICS, source_provenance_table, and source_provenance_manifest documentation included in the deposit. Privacy and ethics The dataset was constructed from pre-existing web-based visual material and did not involve direct interaction with, recruitment of, or intervention involving human participants. The annotation process was limited to person bounding boxes. The dataset does not contain names, identity labels, demographic attributes, behavioral labels, facial-recognition templates, or other intentionally collected biometric identifiers. No attempt was made to determine the identity of any visible individual. The images were selected for evaluating person detection in aerial and top-view scenarios. Close-up imagery and material depicting private or sensitive contexts were avoided during curation. No formal ethics waiver was issued by an institutional ethics body. The research was conducted exclusively for non- commercial academic evaluation, with attention to data minimization, source attribution, privacy protection, and the reuse restrictions imposed by the original content providers.