SAAD-SR: A multi-altitude aerial object detection benchmark evaluated using an altitude-conditioned distillation detector

Object detection in aerial imagery is essential for remote sensing applications such as disaster response, traffic monitoring, and security surveillance. However, detection reliability is often degraded by flight height variation, environmental variability, and reduced object pixel support. These challenges are further compounded by the limitations of existing UAV datasets used for training and evaluation. Although existing UAV datasets have advanced aerial perception research, many lack explicit altitude metadata, provide limited semantic diversity, or contain insufficient real-world variability, thereby restricting the systematic evaluation of altitude-induced scale-regime changes. Consequently, the detector behavior under repeatable flight-height regimes remains insufficiently characterized. To address these challenges, the Scale-Aware Aerial Dataset for Scale Robustness (SAAD–SR) is introduced as a large-scale altitude-stratified benchmark comprising 154K UAV frames, of which 104,255 images are fully annotated for object detection, yielding approximately 1.6M labeled object instances across eight classes under diverse illumination, weather, and scene conditions. The dataset is organized into 15 m, 25 m, and 45 m altitude tiers using synchronized altitude metadata to support altitude-stratified training and evaluation, and the benchmark is evaluated using UAV-oriented and transformer-based detectors. Building on SAAD–SR, the YOLO Scale-Robust (YOLO–SR) framework incorporates the corresponding altitude-regime conditioning value into intermediate features through scalar-conditioned feature modulation and altitude-conditioned teacher–student distillation. Furthermore, a Scale Robustness Index (SRI) is used to quantify detection stability across altitude-induced scale regimes. YOLO–SR achieves a pooled overall validation [email protected] of 0.708 and a mean altitude-wise [email protected] of 0.658, with an SRI of 0.991, while reaching 75.5 FPS in offline evaluation and 13 FPS during onboard UAV operation.