Publicly available object‑tracking models have become the foundation of modern drone‑video sensing because they offer strong generalization, large‑scale training, and reproducible evaluation without requiring custom UAV‑specific architectures. The clearest evidence of this shift comes from the emergence of massive public UAV tracking benchmarks such as WebUAV‑3M, which was released precisely to evaluate and advance deep trackers at scale. WebUAV‑3M contains over 3.3 million frames across 4,500 videos and includes 223 target categories, all densely annotated through a semi‑automatic pipeline 1. What makes this benchmark so influential is that it evaluates 43 publicly available trackers, many of which were originally developed for ground‑based or general computer‑vision tasks rather than UAV‑specific scenarios. These include Siamese‑network trackers, transformer‑based trackers, correlation‑filter trackers, and multimodal variants—models that were never designed for drones but nonetheless perform competitively when applied to aerial scenes. 

The WebUAV‑3M study highlights that publicly available trackers can handle the unique challenges of drone footage—fast motion, small objects; drastic viewpoint changes—when given sufficient data and evaluation structure. The benchmark’s authors emphasize that previous UAV tracking datasets were too small to reveal the “massive power of deep UAV tracking,” and that large‑scale evaluation of existing trackers exposes both their strengths and their failure modes in aerial environments 1. This means that many of the best‑performing models in drone tracking research today are not custom UAV architectures, but adaptations or direct applications of publicly released trackers originally built for general object tracking. 

Earlier work such as UAV123, one of the first widely used aerial tracking benchmarks, also evaluated a broad set of publicly available trackers on 123 fully annotated HD aerial video sequences Springer. The authors compared state‑of‑the‑art trackers from the general vision community—models like KCF, Staple, SRDCF, and SiamFC—and identified which ones transferred best to UAV footage. Their findings showed that even without UAV‑specific training, several publicly available trackers achieved strong performance, especially those with robust appearance modeling and motion‑compensation mechanisms. UAV123 helped establish the norm that drone tracking research should begin with publicly available models before exploring specialized architectures. 

More recent work extends this trend into multimodal tracking. The MM‑UAV dataset introduces a tri‑modal benchmark—RGB, infrared, and event‑based sensing—and provides a baseline multi‑modal tracker built from publicly available components arXiv.org. Although the baseline system introduces new fusion modules, its core tracking logic still relies on publicly released tracking backbones. The authors emphasize that the absence of large‑scale multimodal UAV datasets had previously limited the evaluation of general‑purpose trackers in aerial settings, and that MM‑UAV now enables systematic comparison of publicly available models across challenging conditions such as low illumination, cluttered backgrounds, and rapid motion. 

Taken together, these studies show that the most influential object‑tracking models used in drone video sensing are not bespoke UAV systems but publicly available trackers evaluated and refined through large‑scale UAV benchmarks. WebUAV‑3M demonstrates that general‑purpose deep trackers can scale to millions of aerial frames; UAV123 shows that classical and deep trackers transfer effectively to UAV viewpoints; and MM‑UAV extends this to multimodal sensing. These resources collectively anchor drone‑video analytics in a shared ecosystem of open, reproducible tracking models, enabling researchers and practitioners to extract insights from aerial scenes without building custom trackers from scratch.